Hidden in Plain Sight: Unlocking Actionable Insights in Client Data

• Executive Summary
• 1. Why Precision Matters: The Case for a Different Kind of Evidence
  • 1.1 The Gap inside Most Financial Service Providers' Portfolios
  • 1.2 What Existing Evidence Can, and Cannot, Tell Us
  • 1.3 Why a New Evaluative Approach, and Why PCM
  • 1.4 Precision Causal Modeling: A Conceptual Overview
  • 1.5 How to Read PCM Findings
  • 1.6 What PCM Makes Possible: Use Cases for FSPs and Investors
  • 1.7 A Note on Requirements for PCM
• 2. What the Pilots Revealed: Findings from PCM as an Evaluative Methodology
  • 2.1 The Five Pilots: Design and Context
  • 2.2 A Word on Causality: Why Some Findings are Stronger than Others
  • 2.3 Patterns Emerging from the Pilots
  • 2.4 What Context Determines
  • 2.5 What the Pilots Did Not Resolve
• 3. How PCM Works: A Deeper Look at the Methodology
  • 3.1 Why Conventional Approaches Struggle to Answer the Heterogeneity Question
  • 3.2 PCM's Two Analytical Ambitions
  • 3.3 Interpreting and Using Findings: Levers, Signals, and Indicators
  • 3.4 A Practical Note on PCM's Constraints
  • 3.5 Evaluative and Operational Applications
• 4. Is Your Institution Ready for PCM? A Practical Self-Assessment
  • 4.1 The Five Conditions
  • 4.2 For DFIs and Other Investors
  • 4.3 Indicators of a Productive First Use
  • 4.4 What to Do Now If Your Institution Is Not Yet Ready
• 5. From Evidence to Action
  • 5.1 For FSPs
  • 5.2 For DFIs and Other Investors
  • 5.3 Using Findings Well: Common Errors to Avoid
  • 5.4 Building Sustained Analytical Capacity
• Conclusion: The Case for Precision
• References
• Annex A: Choosing an Evaluation Approach: A Question-Led Guide
• Annex B: The Six Phases of PCM
• Annex C: Five Pilot Partner Profiles
• Annex D: Data Requirements for PCM
• FAQ

Executive Summary

Financial service providers (FSPs) know whether their clients repay. What most do not know is why outcomes can diverge substantially within the same portfolio—in other words, why clients who look similar on paper experience their products very differently—and which specific changes to product design or delivery would improve results for which client groups.

That knowledge gap has consequences.

Given the knowledge gap, institutions cannot reliably know whether their products are working for all of the clients they are designed to serve.

For FSPs, especially those serving low-income and vulnerable populations, it can mean sustaining higher non-performing loan rates than necessary, reducing client retention, and constraining revenue optimization. It can mean making decisions based on a misunderstanding of the core underlying challenge: for instance, designing and redesigning products to enhance client outcomes, when the actual problem is in the product's delivery, ultimately leading to clients being underserved in ways that do not surface as delinquency until it is too late. More fundamentally, it means institutions cannot reliably know whether their products are working for all of the clients they are designed to serve—a question that matters as much for commercially oriented FSPs as for those with an explicit social mission.

For development finance institutions (DFIs) and other investors, portfolio metrics, often an average across investees, can show whether a product worked on balance for the population studied. They are far less able to answer the two questions that sit at the center of due diligence frameworks: who is and is not experiencing positive outcomes, and how confident an investor can be that the product, not pre-existing differences among clients, is the contributing factor (Lahaye et al. 2024). Conventional due diligence asks these questions, but it rarely has the underlying data or methods to answer them with credibility. Without that, portfolio-level assessments of client welfare remain incomplete.

Introducing Precision Analytics and PCM

Precision analytics is the term this guide uses for approaches that combine causal inference with machine learning to reveal what works for whom, rather than what works on average. In particular, the guide introduces Precision Causal Modeling¹ (PCM), a specific method piloted by CGAP in the financial services sector, as a practical methodology for closing FSPs' knowledge gap.

Unlike evaluation methods that focus on examining past performance to reach a verdict, PCM analyzes patterns already present in a portfolio's historical data to surface potential changes that should be made going forward.

PCM is a "target-first" discovery method that combines causal inference with machine learning. It helps FSPs use their own data to identify which product and delivery combinations are associated with better outcomes for specific client segments within their portfolios. Unlike evaluation methods that focus on examining past performance to reach a verdict, PCM analyzes patterns already present in a portfolio's historical data to surface potential changes that should be made going forward. The practical value of this approach lies in using data institutions already hold, such as loan histories, repayment records, and client profiles, rather than requiring new research investments.

The Evidence Gap PCM Addresses

Two decades of randomized controlled trial (RCT)-based research established that average impacts of inclusive credit were modest. But those same studies showed wide distributions of outcomes beneath the average, with gains and losses concentrated among identifiable borrower types. As CGAP's companion Focus Note (Nielsen et al. 2026) on the impact of inclusive credit² documents, heterogeneity is the real signal in the evidence base (Nielsen et al. 2026). CGAP's Impact Pathfinder,³ an interactive evidence platform, makes those heterogeneous findings accessible and actionable at a strategic level. What remained missing was a complementary evaluative methodology capable of generating the same kind of insight at the level of an individual institution's portfolio—one that could inform day-to-day decision-making. Credit and credit-adjacent payment data were the natural domain in which to test such a method: credit is the most high-variance financial service in the evidence base, the one where imprecision causes the most documented harm, and the one where FSPs and financial sector authorities are most likely to hold the longitudinal administrative data a more granular methodology requires.

Testing PCM with Financial Institutions

CGAP partnered with five institutions across four continents to test whether PCM could work in practice for financial institutions: AMK Cambodia, Asociación Grameen Costa Rica, BBVA Microfinance Foundation, FINCA International, and the Banco Central do Brasil.⁴ The pilots ran from late 2024 to 2026 and represent, to the authors' knowledge, the first application of PCM in the financial services sector. Four substantive patterns related to the provision of credit emerged from the pilots with sufficient consistency to serve as structured hypotheses that can help FSPs and other stakeholders in the financial services sector start asking the right questions of their data:

1. Right-sizing loans is critical; in many cases, clients who were underserved by conservative credit ceilings were less likely to achieve positive outcomes.
2. The trajectory of delinquency (i.e., whether a client is recovering or deteriorating) was a far more reliable signal than the mere presence of a missed payment.
3. Structured loan growth of 30 to 50 percent per loan cycle was associated with better outcomes than stagnant or rapidly escalating credit.
4. The quality of the client relationship, including aspects like loan officer tenure and competence of service delivery, can help drive successful outcomes.

The principal near-term opportunity for most institutions is not product innovation; rather, delivery flexibility to adapt offerings to different client segments is critical.

The pilots also confirmed what has long been suspected: there is no one-size-fits-all model when it comes to credit provision. When clients at CGAP's partner institutions got the combination of product features and services that PCM identified as most effective for their particular segment, they often experienced higher rates of successful outcomes. Across the five partners, the delivery gap in outcomes between clients who received the service combinations that the PCM modeling identified as effective for their circumstances and those who did not ranged from 32 to 84 percentage points. These delivery gaps are not differences between institutions, between well-designed and poorly designed products, or even between baseline and endline measurements. They are differences within the same portfolio, between clients who look similar on paper but experienced different outcomes from FSPs' products. These findings highlight that the principal near-term opportunity for most institutions is not product innovation; rather, delivery flexibility to adapt offerings to different client segments is critical. And in most cases, the information needed to build an understanding of where this gap exists and to narrow that gap in outcomes is embedded in an institution's own data; it just needs to be uncovered and acted upon.

What is Needed to Implement PCM

Applying PCM makes three broad demands on an institution:

1. Data: PCM requires longitudinal, client-level records linking baseline characteristics to product exposure and outcomes, and assembling the data is itself substantial work. Across CGAP's five pilots, it consumed between one-third and one-half of total analytical effort.
2. Expertise: The method draws on statistical and causal inference skills that many FSPs do not yet hold in-house, and while external technical partners can fill that gap for a first analysis, internal staff need to understand the method well enough to participate meaningfully in interpreting findings.
3. Effort: Translating segment-level patterns into guidance that loan officers and branch managers can act on is distinct from the analytical work itself and requires deliberate investment.

This guide is frank about each of these requirements throughout. What PCM offers is not certainty, but a different kind of evidence—one that complements and deepens FSPs' current understanding of the average effect of their portfolios and is directly useful for the decisions FSPs and investors need to make.

How to read this guide

This guide is written to help inform FSPs and their funders and investors. It speaks to three audiences: FSP institutional leaders who need to understand what PCM can and cannot do; technical staff responsible for implementing it; and DFIs and other investors considering how to commission or use PCM findings. These different readers will want to navigate the guide differently. Chapter 1 sets out the case for precision analytics, introduces PCM, and lays out its use cases for FSPs and investors. Chapter 2 covers CGAP's five pilot applications, including pilot designs, cross-cutting patterns that emerged, and limitations. Chapter 3 offers a deeper examination of the PCM methodology for more technical audiences. Chapter 4 offers a practical self-assessment for FSPs and investors. Chapter 5 translates the pilot experience into practical guidance. Accompanying technical annexes cover the PCM methodology in greater detail, including the six phases of analysis, data requirements, and PCM's positioning within the broader landscape of evaluation methods.

FSP institutional leaders primarily interested in the business case of PCM—in particular, what it can deliver for a financial institution and whether it is worth pursuing—should focus on Chapters 1, 2, 4, and 5. Readers primarily interested in understanding the methodology—how PCM works and how its claims should be interpreted—should focus on Chapter 3 and Annexes A and B. DFIs and other investors should focus on Chapters 1, 2, 4, and 5.

For institutional leaders, the core message of this guide is that the information needed to improve client outcomes—to match product design and delivery to what works for each client segment—is, in many cases, already embedded in an institution's own data. PCM makes it visible and actionable. For DFIs and other investors, the core message is that the gap between headline portfolio metrics and actual client welfare is larger and more addressable than conventional due diligence typically reveals, not because due diligence does not ask the right questions but because it rarely has the data or methods to answer them with confidence. For technical staff, the guide provides the detail needed to get started with PCM.

1. Why Precision Matters: The Case for a Different Kind of Evidence

1.1 The Gap inside Most Financial Service Providers' Portfolios

Most financial service providers (FSPs) know whether their clients repay. What very few can determine is why outcomes diverge substantially within the same portfolio—why clients who look similar on paper experience their products very differently—or which specific changes to product design or delivery would improve results for which client groups. This knowledge gap has direct operational consequences: it can mean sustaining higher non-performing loan rates than necessary, reducing client retention, and constraining institutions' ability to optimize revenue. More fundamentally, it means institutions cannot reliably deliver on the social mission aspect of their work. The question is how to bridge this knowledge gap.

Routine reporting tracks outreach and repayment. Some FSPs have invested in client surveys or evaluations that give them reasonable confidence that, on average, their products improve clients' lives. Some can even see from disaggregated data that benefits vary across their client base—in other words, that some groups fare better than others. But knowing that rural clients fare worse on average does not tell an institution which rural clients, with which loan configurations, under which delivery conditions account for this pattern. Yet these are exactly the distinctions that matter for day-to-day decisions. How products are priced, how loan ceilings are set, how repayments are structured, and which clients to target all depend on understanding that finer-grained variation: who responds well to what.

Clients poorly matched to a product may take on debt they cannot put to productive use, experience repayment stress, and leave the portfolio no better off, or worse, than when they first engaged with the institution.

Building this understanding is also a social mission imperative. Treating clients with very different circumstances as equivalent and able to be served by the same product on the same terms is not just analytically imprecise, it limits FSPs' ability to deliver positive outcomes. Clients poorly matched to a product may take on debt they cannot put to productive use, experience repayment stress, and leave the portfolio no better off, or worse, than when they first engaged with the institution.

This guide makes the case for a new approach—one that can complement existing methodologies—to generate this level of insight. It introduces precision analytics, a family of approaches that combine causal inference with machine learning to move beyond what works on average and reveal what works for whom, under which conditions. Rather than summarizing effects across a population, precision analytics surfaces variation within it, making that variation visible, interpretable, and actionable. Specifically, this guide introduces Precision Causal Modeling (PCM) as a practical methodology for closing this knowledge gap. PCM is a "target-first" discovery method that can help FSPs use their own data to identify which product and delivery combinations are associated with better outcomes for specific client segments within their portfolios by combining causal inference with machine learning. The sections that follow explain the evidence gap PCM addresses, what CGAP's pilots found, what PCM is, and what it can do in practice for FSPs and investors.

1.2 What Existing Evidence Can, and Cannot, Tell Us

A series of large randomized controlled trials (RCTs) published in the 2010s generally found that the average effects of microcredit on income, poverty, and women's empowerment were modest, often statistically indistinguishable from zero (Augsburg et al. 2015; Banerjee et al. 2015; Angelucci, Karlan, and Zinman 2015; Attanasio et al. 2015; Crépon et al. 2015; Tarozzi, Desai, and Johnson 2015; Karlan and Zinman 2011). However, averages can mask important heterogeneity. Several of these studies also found that effects varied across groups of borrowers. Additionally, Rachael Meager's two Bayesian hierarchical analyses (2019, 2022) of the same seven studies similarly found this pattern: even where average effects were modest overall, households with prior business experience often saw larger and more heterogeneous effects than other borrowers—a subgroup signal that complicates the standard 'modest average impacts' finding from first-generation RCT evidence. But building a deeper understanding of this variation—especially at the individual FSP level—requires evidence at a level of precision that average-effects analysis typically does not provide.

RCTs remain the strongest method for estimating average causal effects and provide important insights to FSPs and their investors. However, traditional survey-based RCTs, which remain the modal approach in the financial inclusion sector, are often costly and slow,⁵ and they are designed principally to produce an average estimate rather than to map the detailed distribution of effects across client types. Quasi-experimental evaluations—methods such as difference-in-differences, regression discontinuity, and instrumental variables—can produce credible causal estimates when randomization is not feasible but rely heavily on certain assumptions and are ultimately designed to estimate average treatment effects. Routine monitoring provides essential operational information but cannot answer causal questions. Predictive analytics can anticipate which clients will default or take up a product, but they identify characteristics of clients who tend to succeed and depend on correlations rather than attempting to identify causal relationships.

What has been missing is a complementary evaluative methodology capable of generating granular, institution-specific, actionable insight—one that works from data FSPs already hold and that is oriented toward operational decision-making rather than retrospective verdicts. Methods to measure heterogeneous treatment effects have been evolving in recent years, including the expansion of machine-learning-based options. PCM, the method tested by CGAP, provides one such option for institutions interested in delving into the question "what works, for whom, in what circumstances". It is a commercial methodology developed and owned by BCT Partners. While pilot applications for financial institutions are documented in this guide, the core methodology has not, at the time of writing, been subject to independent peer-reviewed publication. The methodological foundations should be regarded as promising and practically tested rather than as having attained the standing of established, peer-validated scientific practice. Alternative observational causal inference approaches that have been peer-reviewed—including propensity score matching, regression discontinuity designs, and doubly-robust estimators—exist and may merit consideration alongside or instead of PCM in specific contexts (see Annex A).

1.3 Why a New Evaluative Approach, and Why PCM

It was partly Meager's finding—that beneath a moderate average effect, one identifiable group of borrowers (those with prior business experience) showed substantially larger and more heterogeneous effects than others—that prompted CGAP to look for a methodology capable of locating that kind of segment-level variation within an individual institution's portfolio. Among several approaches considered, PCM emerged as the strongest candidate, capable of working with data that most FSPs already hold and of generating findings that could directly inform the product design, targeting, and delivery decisions that determine outcomes at the client level.

CGAP partnered with five institutions across four continents to test whether PCM could work in practice: AMK Cambodia, Asociación Grameen Costa Rica, BBVA Microfinance Foundation, FINCA International, and the Banco Central do Brasil (BCB).⁶ While, to the best of CGAP's knowledge, this was the first time PCM has been used in financial services, the methodology has previously been applied in sectors such as workforce development, education, and mental health to enable customized case management and enhance outcomes for clients (see York 2024); it was PCM's capacity to enable tailored service delivery with the intention of producing better client-level outcomes that CGAP sought to examine in the context of financial services.

Between late 2024 and 2026, each partner worked with CGAP to apply PCM to its own administrative data (e.g., loan histories, repayment records, and client profiles). In two cases, the administrative data were supplemented by available client survey data. Each pilot asked a version of the same underlying questions: which clients are receiving the combinations of services associated with positive outcomes for people in similar circumstances; which are not; and what difference would it make if they received the optimal combination for their client profile?

For these pilots, credit and credit-adjacent data were the natural starting domain. Credit is the most high-variance financial service in the evidence base, and it is the financial service where imprecision causes the most documented harm and where FSPs and financial authorities are most likely to hold the longitudinal, client-level administrative data a granular methodology requires. (See Box 1 for more on why credit.)

CGAP's five partners brought significant diversity to the pilots. The four FSPs vary in size, geography, and lending approach. The BCB brought a regulator's perspective and a different kind of data, drawn from the Pix instant payment platform, which enables users to send or receive payment transfers almost instantaneously at any time (Banco Central do Brasil n.d.). The type and quality of data available differed substantially across partners, and as Chapter 2 explains, so does the strength of causal claims that resulted. Nevertheless, the findings from these five pilots were consistent enough across very different institutions and geographies to suggest patterns worth taking seriously (see Box 2).

1.4 Precision Causal Modeling: A Conceptual Overview

PCM is a "target-first" discovery method that combines causal inference with machine learning. Rather than starting with a predefined product and asking whether it worked, PCM reverses that logic. It begins with the outcome the institution cares about most (such as poverty reduction, financial resilience, business growth) and works backward through historical administrative data to discover which specific combinations of product features and delivery approaches are most strongly associated with clients achieving that outcome (see Box 3 for more details). It does so by identifying comparable segments ("Block Matched Groups," see Key Terms) to discover which combination of product features and service approaches are associated with stronger outcomes for each matched group. It then tests whether those associations hold under causal scrutiny—examining how comparable the matched groups actually are and how sensitive the findings are to unobserved differences between them—to assess how much confidence the patterns warrant.⁷ (See Figure 1 for an overview of PCM.)

The practical value is that it uses data institutions already hold—such as loan histories, repayment records, and client profiles—rather than requiring new research investments. In portfolios where product configurations vary across loan officers, branches, time periods, or policy changes, PCM makes that variation visible and identifies what it means for client outcomes.

Figure 1 – An Overview of PCM

1.5 How to Read PCM Findings

PCM findings are not prescriptive. They are structured, directional hypotheses.

PCM findings are not prescriptive. They are structured, directional hypotheses: patterns observed within an institution's data with enough specificity to be useful as a starting point for institutional inquiry. Causal claims from PCM are not equivalent to those from a well-executed RCT, and it is important to note that the broader empirical literature on observational causal inference reveals a persistent challenge: estimates derived from observational methods frequently diverge from those obtained by running the same comparison as an RCT, even when the observational method is carefully implemented and the sample is large. This is an intrinsic limitation of any selection-on-observables approach, including PCM, and it reinforces the Guide's consistent advice to treat PCM findings as structured, directional hypotheses to be investigated and piloted rather than as verified causal facts. Nevertheless, PCM findings are substantially more informative than simple correlational analysis and can be more actionable for institutional decision‑making than conventional descriptive or average‑effect approaches.

PCM findings fall into three categories that warrant different responses:

• Actionable levers are product design features, policies, or practices that the institution can modify, such as loan progression structures, delinquency tolerance mechanisms, and loan purpose flexibility.
• Assessment signals are client characteristics that help identify capacity, risk, or trajectory; these belong in intake and targeting decisions, not in product design.
• Milestone indicators are variables that track whether a client is on a successful trajectory. They are useful for monitoring but should not be treated as levers.

Methodologically, PCM is best understood as a discovery tool and a precursor to more targeted testing rather than a substitute for experimental methods; it is particularly well-suited to situations where an institution has accumulated variation in its data and does not yet know what to make of it. PCM's role is to bring structure to that variation and identify the most promising hypotheses. If an institution already knows which specific product feature it wants to evaluate and can randomize delivery, a prospective experimental design is the right tool. PCM is for the earlier phase: discovering what is worth testing in the first place.

1.6 What PCM Makes Possible: Use Cases for FSPs and Investors

Precision analytics opens a new set of capabilities for FSPs. The applications in Table 1 illustrate what becomes possible when institutions use precision analytics like PCM to ask different questions of data they already hold. These use cases are organized by audience and by whether the primary value is impact-related or commercial.

TABLE 1. PCM Use Cases

Source: Authors.

These applications share a common logic: each becomes possible when an institution can move from asking what works in general to asking what works for whom, under which conditions. That shift—from aggregate to segmented, from descriptive towards causally informed, from static to dynamic—is what precision analytics enables.

For FSPs, the payoff is twofold. The commercial return is clear: the potential for lower non-performing loan rates, stronger client retention, and more targeted product decisions. The mission return is equally direct: fewer clients served by product feature and service combinations that do not work for them, and a clearer line of sight from institutional practice to client welfare.

1.7 A Note on Requirements for PCM

PCM is not a lightweight exercise. The methodology requires longitudinal, client-level data linking baseline characteristics to product exposure and transformative outcomes. PCM can leverage data institutions already hold, provided that the data captures some level of product exposure and client outcomes (or proxy outcomes), rather than requiring new data collection. However, preparing that data for analysis is demanding. Across CGAP's five pilots, data preparation consumed between one-third and one-half of total analytical effort before modeling could begin.¹⁰

PCM also requires statistical expertise that most FSPs do not yet hold in-house. CGAP's pilots were all conducted with external technical partners, and that model works well for a first use of PCM. The more important prerequisite is that internal staff understand the method well enough to participate meaningfully in the interpretation phase: findings that only an external technical partner understands will not change operational practice.

Further, translating findings into guidance that loan officers and branch managers can act on requires deliberate effort and investment, separate from the analytical work itself. A first PCM analysis commonly identifies where operational refinement is most warranted; developing the tools and processes to act on those findings is a subsequent step.

Finally, PCM as a one-time exercise has limited institutional value. That value multiplies when the analysis is repeated, findings accumulate, and each cycle builds on the last. Chapter 4 provides a practical self-assessment for FSPs and investors considering whether and how to proceed.

2. What the Pilots Revealed: Findings from PCM as an Evaluative Methodology

The five pilots described in this chapter were designed to test whether PCM could work in practice for FSPs. CGAP wanted to explore whether the methodology could be applied to real institutional data, in real operational contexts, and produce findings specific enough to be useful. That is a distinctly different question from whether PCM works as an embedded, ongoing decision-support tool integrated into an institution's analytical processes. As such, the pilots were evaluative exercises: each was conducted as a discrete analysis, drawing on historical administrative data, with findings reported at the end of a defined analytical period. "Evaluative" here should not be read as retrospective or terminal. The intent throughout was to surface patterns actionable enough to change what institutions do next. However, they were not designed to test PCM as a continuous feedback loop between analysis and operational decision-making, though FINCA International's extension of the pilot work (see Box 4) offers the earliest indication that such a transition is achievable.

What the pilots establish is a foundation: evidence that the methodology is viable, that it surfaces findings with operational implications, and that those findings are consistent enough across very different institutional contexts to be taken seriously. The question of how PCM performs when embedded as a recurring institutional practice is the next frontier.

2.1 The Five Pilots: Design and Context

The five partner institutions were selected to participate in these pilots based on their willingness to explore heterogeneous impact within their portfolios, availability of longitudinal client data, and data that captured client-level outcomes. They differed substantially in geography, lending methodology, data availability and environment, client population, and target outcome. AMK Cambodia combined survey and management information system (MIS) data with national poverty indicators and geographic context data from a national database. Asociación Grameen Costa Rica relied entirely on administrative records from a group-lending institution exclusively serving women. BBVA Microfinance Foundation's data covered multiple Latin American countries, with separate analysis of group and individual lending. FINCA International's analysis spanned various countries in Africa and Central Asia using two rounds of client survey data alongside administrative data. The BCB cross-referenced Pix payment outcomes with credit information using administrative data only. The breadth of this variation is what makes the cross-pilot synthesis meaningful, and what gives the consistent patterns across it their practical weight. (Profiles of each partner institution are provided in Annex C.)

The CGAP pilots leveraged two variations of PCM analysis: episode-based implementation and window-based implementation. In the first, episode based implementation, each client's journey is treated as a bounded episode: analysis begins from the client's first engagement with the program (in this case, their first loan) and follows them through to a defined endpoint.¹¹ Three pilots used this approach. In the second, window-based implementation, analysis is anchored to a fixed slice of time. The key difference is that the window based approach accounts for both the client's starting characteristics and their accumulated experience with the FSP at the start of the window (i.e., the baseline). The analysis then examines what happens as clients experience a product between window start and window end. Two pilots used this approach. Ultimately, the choice of study time period should be made jointly with the choice of target outcome at the start of the analysis, grounded in the data available and how clients' journeys are structured in practice. (See Table 2 for more details.)

TABLE 2. Summary of Pilot Designs Across the Five Pilots

Source: Authors.

An additional design choice is worth noting: whether to define client segments directly from the data or to rely on the institution's existing segmentation frameworks. While data-derived segments can be analytically powerful, they may not lead to insights that can be easily operationalized. For instance, the initial AMK analysis produced seven data-derived segments, with each segment showing dramatically different success rates of clients being above the national poverty threshold. But what the data can reveal and what staff can act on are not the same thing. The data-derived segments for AMK were built from community-level contextual variables drawn from a national database. These data do not appear in AMK's operational data, which made it impossible for staff to know which segment a client would fall into when working with them. A second model using AMK's existing livelihood-based client personas also produced meaningful insights into the product features and delivery approaches that were most effective in driving positive outcomes for each client persona. The second model was also more actionable, as loan officers had this livelihood-based classification at their fingertips.

The lesson here is not that pre-existing frameworks are always preferable—the other four pilots leveraged data-driven segmentation to identify actionable findings—but that the test for any segmentation approach is whether the people making decisions about clients can readily use it. Institutions considering PCM should review their existing categorization systems early. In some cases, a well-designed institutional typology may be preferred over a data-derived alternative simply because it is already embedded in how staff think and what systems record.

2.2 A Word on Causality: Why Some Findings are Stronger than Others

As noted in Section 1.5, PCM findings are not prescriptive. While PCM seeks to identify causal patterns, the strength of this causality varies. Ultimately, PCM leverages observational data and depends on an institution's data (e.g., available data variables, portfolio coverage, data quality, etc.). Causal credibility is pursued, not guaranteed, by identifying comparable matched groups within the data set. Within each group, PCM checks whether clients who received the experience patterns associated with success differ systematically from clients who did not. When the two groups look similar on observable characteristics available in the data (in other words, when there are strong signals that the comparison is "apples to apples"), the findings carry stronger causal credibility. Where the groups differ in ways that could also explain the outcome, the findings remain valuable as associational evidence that warrants further investigation, but translating them into operational changes requires people who understand the institution, the clients, and the local context; further testing may also be warranted before changes are rolled out.

The strength of causal claims within CGAP's pilots differed across institutions and for multi-country analyses, across countries depending on the data, sample size, and data quality available. This variation in causal credibility is not a weakness of the method. This variation reflects what can be inferred from observational data available in different institutional contexts. PCM does not eliminate confounding, which occurs when the relationship between a treatment (such as a loan) and an outcome is distorted by a third variable that affects both (see Key Terms). Rather, PCM reduces confounding systematically and makes the remaining assumptions explicit. Specifically, PCM addresses confounding from observed characteristics; residual bias from unobserved factors cannot be excluded, only assessed.

Institutions and investors should therefore expect findings to fall along a spectrum of causal confidence. Analytical partners should be equally explicit about where each finding sits on that spectrum. This spectrum extends to Section 2.3's presentation of results: the specific percentages reported in Section 2.3 are point estimates, and readers should be aware that they carry uncertainty that is not always made explicit in the narrative. The ranges (e.g., 56–97 percentage points) reflect variation across different matched groups. None of the pilots produced formal confidence intervals for these figures, so all point estimates should be read as indicative of direction and approximate magnitude rather than as precise measurements. The findings presented in Section 2.3 should be read with this in mind.

2.3 Patterns Emerging from the Pilots

No single loan decision drove success. It was the cumulative trajectory that mattered.

Across the five pilots, four substantive patterns in the provision of credit appeared with enough consistency to serve as a diagnostic framework for any institution examining its own data. These findings derive from different data availability, data quality, and target outcomes, and as such, they are structured hypotheses: patterns observed in enough contexts, with enough specificity, to be useful as a starting point for asking the right questions of an institution's own data. Where a finding holds across multiple independent contexts with well-balanced matched groups, the language reflects that stronger position. Where evidence is thinner, the language is deliberately more tentative. Additionally, considerations of survivorship bias are noted.

Under-Lending, Not Over-Lending, Was the Most Consistent Predictor of Failure

Across the pilots, clients who were under-served by conservative credit ceilings achieved target outcomes at rates systematically lower than clients receiving appropriately sized loans. Importantly, though, while "right-sizing" loans was critical, it did not always mean that more was better: what counts as "right-sizing" can look different for different clients, depending on institutional model and context. In standardized lending models that follow a set loan progression, the key factor that an institution can adjust to influence outcomes is progression velocity: how quickly the institution advances proven clients to adequate capital levels. In individualized models, the key factor is initial assessment quality, in other words, the accuracy of the first loan decision. Across both models, the pilots identified four consistent parameters:

1. Minimum loan floors must exceed subsistence-level amounts to enable productive investment, not just cover basic needs.
2. Progressive increases of around 30 to 50 percent per loan cycle are associated with better outcomes than either stagnant loan sizes or large, rapid increases.
3. Household-level debt burden must be assessed alongside individual loan size.
4. Clients with irregular income need right-sized, purpose-focused loans, not maximum possible credit.

Together, these findings suggest that conservative credit limits, intended to manage risk, may in practice compound that risk. For instance, in one partner institution's results, the gap in success rates between clients in the lowest and highest loan-size tiers ranged from 56 to 97 percentage points across the different matched groups, a delivery gap consistent with the pattern seen across all five pilots. At another partner institution, successful clients received larger first loan disbursements in 30 of 34 segments analyzed, with the advantage increasing with client capacity: For clients with the strongest baseline business indicators (measured by business surplus, total assets, or business equity, depending on the country), successful borrowers started with loans 76 percent larger on average than their unsuccessful counterparts and ended the study period with loans 2.7 times the size of unsuccessful clients' final loans.

Another partner institution's analysis found that most client segments showed a clear pattern: Larger loans were associated with successful outcomes. For instance, successful agriculture clients had loans approximately three times as large than unsuccessful ones in the same matched group. However, there was one exception to this within the institution's portfolio. Casual labor clients showed the opposite pattern: successful clients had smaller loan amounts but adequately sized disbursements that were well-matched to clients' needs. They succeeded when loans were right sized for a focused purpose, not when they took on larger, multi-purpose debt. This demonstrates that for this segment, "right-sizing" meant matching capital to purpose, not simply increasing the amount.

Finally, one partner institution provided the clearest qualification to the "more is better" pattern that seems to emerge from the other four pilots. Across eleven countries studied in this pilot, average loan size appeared among the top drivers of outcomes in six countries. However, the direction of the relationship varied by matched group. In some matched groups (e.g., established rural borrowers working in the trade, service, and agriculture sectors), clients who received the effective configuration had larger loans; in other matched groups, they had smaller ones. There was no universal optimal loan size. This is the strongest evidence across the pilots that appropriate sizing depends on segment characteristics: client type, savings behavior, prior lending history, and business sector.

A natural question raised by this pattern is whether it reflects survivorship rather than an effect of loan sizing: Clients who remain in a lending relationship long enough to receive larger or more numerous loans may simply be the ones who were already succeeding, rather than succeeding because of the loan size itself. The window-based implementation of PCM used in two of CGAP's pilots was designed to mitigate one form of this risk by retaining all clients active during the observation period—including those with only one loan and those with unresolved outcomes—rather than restricting analysis to clients who reached a minimum number of loan cycles. It does not fully eliminate the concern, however, since loan count and cumulative size are still measured over a fixed window and can partly reflect how long a client has been observed rather than the sizing decision alone. Future applications of PCM can explore this more directly by treating loan-by-loan progression as an explicit sequential process, tracking attrition rates at each stage between clients who received larger versus smaller loans, and including tenure or time-in-program as a formal confounder in the causal credibility diagnostics. Where sample size allows, comparing outcomes among clients with similar observed tenure would further separate the effect of loan sizing from the effect of simply remaining in the program longer.

Delinquency Trajectory Was a More Reliable Signal Than Delinquency Occurrence

Across every partner institution, rigid responses to delinquency that treated all missed payments equivalently were associated with worse client outcomes. The key distinction is between a deteriorating repayment pattern and a single missed payment. The former is a real warning signal; the latter may reflect nothing more than irregular but manageable income flows. While standard portfolio monitoring does not make this distinction, PCM can. The overarching finding from the pilots is not that "delinquency is fine", but that whether delinquency predicts failure depends on three things: income predictability and livelihood type, whether the client is recovering or deteriorating, and the lending methodology and institutional infrastructure around them. This finding indicates that institutions should monitor delinquency trajectory rather than enforcing rigid payment schedules, and should calibrate tolerance to the client's income profile.

Importantly, the pilots found that the mechanism through which delinquency tolerance operates varied substantially by context. For example, clients at one partner institution who could adjust their payments when they were late were much more likely to succeed. Depending on the matched group, success rates were 31 to 72 percentage points higher. For example, in one matched group, 85 percent of clients who adjusted their payments frequently achieved positive outcomes, compared to just 22 percent of those who rarely did.

At another partner institution, two distinct country-level patterns emerged: in three countries, successful clients consistently showed lower maximum days past due; in two other countries, they did not. This split likely reflects differences in credit culture and enforcement norms: In countries where on-time repayment expectations are strong and consistently enforced, falling behind reliably signals client distress, while in countries where payment timing norms are more flexible, moderate delinquency may reflect local customs rather than financial difficulty.¹⁴ Across all five countries included in this analysis, successful clients carried higher average irregular balances in absolute terms, not because of worse repayment, but because their larger loans meant even modest delinquency produced higher arrears figures.

Another partner institution found that livelihood type is a key moderator: agriculture clients showed repayment variability nearly three times as high among successful versus unsuccessful borrowers, driven by irregular but manageable income patterns, not financial distress.

For another partner institution, near-zero delinquency during the study period emerged as a consistent differentiator between clients who received the most effective combination of services and those who did not. Delinquency prevention through appropriate loan sizing, payment structure, and early intervention emerged alongside the use of savings products and customer care quality¹⁵ as key drivers of positive outcomes. This frames delinquency management as a program quality metric, not merely a portfolio risk metric.

Growth Velocity and Progressive Structuring Were Consistently Associated with Success

Across all five pilots, clients who achieved positive outcomes tended to be those on structured progression pathways, characterized by growing loan sizes, improving payment capacity, diversifying loan purposes, and deeper institutional relationships over time. No single loan decision drove success. It was the cumulative trajectory that mattered.

For example, in one partner institution's analysis, growth velocity was the strongest universal predictor of successful outcomes across multiple countries, with successful clients' final loans approximately twice the size of unsuccessful clients' final loans. In one particular country, unsuccessful clients in higher-capacity segments showed negative growth trajectories; their loan portfolios actually contracted over time. The critical distinction was not between fast and slow growth, but between growth and contraction.

At another partner institution, concentrated capital allocation to primary loan purposes consistently outperformed dispersed small allocations across all client segments. The specific thresholds varied but the pattern of concentrated versus fragmented investment remained constant. Progressive loan sizing functioned as "financial infrastructure" that substituted for missing market opportunities in constrained contexts.

An important caveat to this pattern must be noted: clients identified as following structured progressive pathways are by definition those who continued borrowing over time.

Borrowers who dropped out or experienced distress—precisely those most likely to have the worst outcomes (e.g., defaults, distress exits)—are absent from the data used for analysis, and this survivorship bias may be a significant driver of the pattern. These findings should therefore be read as patterns observed among continuing borrowers, rather than as evidence that progressive pathways cause positive outcomes for all who begin them. The window-based approach used by two pilots partially addresses this by requiring only that clients be present at both window start and end, but dropouts between periods remain outside the analysis.

How Services Were Delivered Mattered as Much as What Was Delivered

Two partners had data that allowed the analysis to look beyond product terms¹⁶ to the delivery experience, and both cases surfaced findings suggesting that the quality of the relationship between FSPs and clients is not separate from the product; it is part of it.

In one partner institution, loan officer tenure of nine or more years generated up to 44-percentage point improvements in outcomes for the most vulnerable client segments. However, for clients who were graduation-ready, newer officers outperformed veterans. Accumulated microfinance expertise appears to be most valuable for clients who are still developing and can become a constraint for those ready for commercial-scale products.¹⁷

Additionally, for this partner, center¹⁸ effectiveness¹⁹ was similarly powerful and consistent for certain client groups. For instance, center effectiveness delivered 15 to 20 percentage point improvements for individuals living in districts with universal internet connectivity and moderate infrastructure access and 25 to 45 percentage point improvements for vulnerable populations lacking infrastructure.

For the other partner institution with delivery experience data, customer service quality, measured by staff rapport, competence, service speed, and fair treatment, emerged as a leading driver of outcomes in most countries included in the analysis. However, loan product experience—client satisfaction with term flexibility, processing time, application process, and documentation requirements—was the most frequent driver category, outranking customer care in frequency (if not always in magnitude). The implication is that product accessibility and administrative experience are not secondary to relationship quality; for many client segments, they are the primary interface through which the institution either builds or erodes client progress.

These findings could not be tested across all five partners because the relevant data was not available for all of them, and as such, they are not cross-cutting patterns. But the fact that two partners, working with different data in different contexts, independently surfaced delivery experience as a powerful predictor of positive outcomes is a useful signal, although not immune to confounding from other variables. For institutions that track loan officer assignments, group membership, or client satisfaction, this is testable in their own data. For institutions that do not currently collect this kind of information, a practical first step is to start collecting this data and analyzing it to see whether there are areas where outcomes vary and then work out what the high performers are doing differently.

2.4 What Context Determines

No single lever is universally the most important.

Perhaps the most important cross-cutting finding from CGAP's pilots is that the ranking of institutional decisions or design features that an institution can change through policy or product design—in other words, the "levers" that an institution can control—varies by context. No single lever is universally the most important. What varies is not the direction of the findings, but which specific lever has the greatest marginal impact for a given client segment.

The pilots identified five moderators that appeared consistently across all five pilots and informed which of the levers described above (i.e., delinquency tolerance, loan sizing, progression velocity) mattered most for a given client segment. Table 3 summarizes these moderators. None of these is particularly surprising; all are familiar dimensions of credit program design. What is new is that PCM shows precisely which one matters most for a specific client in a specific circumstance, a level of precision conventional analysis cannot offer. The same institution, applying the same set of products, can find that delinquency tolerance is the critical lever for one segment, and loan sizing is the critical lever for another.

TABLE 3. Cross-Pilot Findings and Their Moderating Factors

Source: Authors.

These levers do not operate in isolation. Their relative importance depends on the intersection of institutional model, client livelihood type, and credit culture.

2.5 What the Pilots Did Not Resolve

Clarity about what the pilots did not resolve is as important as the findings they produced. Three significant open questions remain.

The first concerns the strength of causal claims. PCM generated credible, actionable evidence across all five pilots, but not findings with the confidence of a well-executed RCT. The causal claims varied in strength depending on how well-balanced the matched groups were and how comprehensively the available data captured the factors that drove selection into different product configurations. That matters for how findings should be used. As noted in Section 2.2 above, they should be read as directional evidence warranting operational investigation, not as verified causal estimates warranting immediate, unqualified policy change.

The second is the question flagged in the opening of this chapter: whether PCM works as an embedded, ongoing decision-support instrument, not just as a periodic evaluative exercise. CGAP's pilots were designed to test its feasibility as an evaluative exercise; they were discrete analyses conducted over a defined period, with findings delivered at the end. They were not designed to test what happens when PCM outputs feed continuously into loan officer guidance, product recalibration, or client targeting decisions, or whether the findings that emerge from embedded, real-time use differ meaningfully from those produced by a retrospective analysis of the available data. FINCA International went the furthest and began to extend the pilot work (see Box 4), offering the earliest indication that the transition from evaluative to embedded use is achievable, but it does not constitute a test of embedded PCM in full operation. This gap is not a shortcoming of the pilots: testing PCM's evaluative use for financial institutions was the right starting point. Rather, it is an acknowledgment of where the evidence currently stops and where future research is needed.

The third is durability: translating PCM from an externally supported exercise to an analytical capability that FSPs own, repeat, and build over time. In every pilot, analysis was conducted with external technical support. FINCA International tested elements of in-house replication and demonstrated that durable internal capacity is achievable with the right resources and institutional commitment, but it remains the exception. The question of what it takes for an FSP to run PCM independently, repeatedly, and without sustained external scaffolding—in terms of both analytical capability and resources—was not resolved. PCM as a one-time exercise has limited institutional value; that value multiplies when the analysis is repeated, findings accumulate, and each cycle builds on the last. Achieving this requires investment in data systems, analytical roles, and decision-making processes.

3. How PCM Works: A Deeper Look at the Methodology

There is a difference between a product that reaches people who tend to succeed and a product that causes people to succeed.

There is a difference between a product that reaches people who tend to succeed and a product that causes people to succeed. For institutions serious about improving client outcomes, whether for commercial or social-mission reasons, that distinction matters. This chapter explains, in more depth than Section 1.4, the analytical approach that PCM uses to make that distinction and what it can and cannot tell institutions about their products and clients.

3.1 Why Conventional Approaches Struggle to Answer the Heterogeneity Question

The fundamental difficulty in estimating cause and effect from institutional data is something researchers call "selection bias": clients are not randomly assigned to products. Clients who take up a particular loan product may differ systematically from those who do not in age, gender, prior loan history, business type, risk tolerance, and other characteristics. If we simply compare the outcomes of the two groups, we cannot separate the product's effect from the fact that the groups were already different.

Predictive tools, including standard regression models, tell institutions which clients are likely to succeed, but not what caused their success. Confusing prediction with causation leads to decisions that follow success rather than create it.

Even where a credible average causal effect can be estimated, it may be of limited operational use. An average tells institutions something about overall performance, but not how to improve it, whom to target, or how to adapt the product to different circumstances. Standard survey-based approaches to understanding variation split clients on one or two characteristics, such as gender or business ownership. They cannot map how effects vary across the many interacting factors that jointly determine whether a product helps or harms a given client.

3.2 PCM's Two Analytical Ambitions

Section 1.4 introduced PCM's two analytical ambitions at a high level. This section examines each in more technical depth, starting with how PCM seeks to build causal credibility. This means moving beyond simple correlation toward a more credible basis for asking whether differences in product experiences are associated with differences in outcomes, and whether those associations are likely to reflect background differences between clients.

PCM integrates two major causal inference traditions to achieve these ambitions. First, PCM analysts and FSP staff articulate an expected causal pathway and accordingly tag program data into five categories (see Box 5) before any machine learning modeling is conducted. This tagging enables the subsequent modeling to identify how context and baseline data shape both response to the program and outcomes, to prevent confounding pathways (Pearl 2009),²¹ and to separate what the program did from what would have happened due to background conditions. Second, PCM uses matching methods and related techniques that create credible comparison groups from within the existing client base.²² Among clients who were equally plausible candidates for a given product feature, whether a particular client actually received that feature may be close to random, which allows PCM to generate a credible approximation of the causal effect without requiring a formal experiment.²³ The critical assumption is that all relevant differences between treated and untreated clients are captured in the observable data. This assumption is seldom fully satisfied, and PCM cannot establish causal effects with the same confidence as a well-executed RCT. Unobservable characteristics—i.e., factors not captured in institutional data—remain a persistent challenge for any selection-on-observables approach. The guide's consistent advice to treat findings as structured directional hypotheses rather than verified causal facts reflects this limitation. Instead, PCM offers credible, practically useful estimates that can guide operational decisions and help sharpen the specific treatment patterns institutions may later choose to test through RCTs or other prospective designs.

The second ambition is precision: to move beyond averages and understand how the effect of a service pattern differs within groups of comparable clients. After forming groups of clients who start in similar circumstances (BMGs), PCM uses machine learning algorithms to test whether particular combinations of product features and service approaches are associated with significantly higher rates of target outcome achievement within each group. In other words, rather than assuming the same configuration works for everyone, PCM identifies which specific experience pattern is associated with better results for each distinct type of client.

PCM is best understood as a hybrid: causal inference for its identification strategy,²⁴ machine learning for its capacity to detect heterogeneity. It is the combination that allows PCM to generate evidence—from an institution's own data—which is both plausibly causal and precise enough to inform concrete action. PCM was developed and tested by BCT Partners, in consultation with government and private funders, to help address complex social issues through the use of large-scale data (York 2024).

PCM operationalizes its two analytical ambitions through six sequential phases:

1. Agreeing on a theory of change and target outcome
2. Assembling and assessing the quality of available data
3. Constructing comparable BMGs
4. Estimating heterogeneous treatment effects using random forests and SHapley Additive exPlanations (SHAP)
5. Interpreting and translating findings for operational use
6. Conducting an equity analysis

Annex B provides a detailed account of each phase for analysts and technical staff, and Annex D outlines key data requirements.

3.3 Interpreting and Using Findings: Levers, Signals, and Indicators

A recurring question across CGAP's pilots was how to understand and use PCM findings: do the patterns identified represent the causes of client success, or do they identify characteristics of clients who were already on a successful trajectory? The answer depends on which specific finding is under consideration, and that distinction materially affects how findings should be used.

PCM findings fall into three categories, which warrant different institutional responses.

Actionable levers are product design features, policies, or practices that the institution can modify, such as loan progression structures, delinquency tolerance mechanisms, and loan purpose flexibility. Where PCM identifies that a lever is associated with better outcomes for certain segments and the evidence is sufficiently strong to warrant action, the institution should consider a policy or operational change. Any adjustment should be assessed against portfolio sustainability and cost-to-serve implications before implementation, and changes that survive this test warrant controlled piloting before full rollout.

Assessment signals are client characteristics or behaviors that help institutions identify capacity, risk, or trajectory, rather than what the institution can change (e.g., existing business assets, market access conditions, and community-level economic context). These warrant use in client assessment to inform better product targeting, not in program design.

Milestone indicators are variables that track whether a client is on a successful trajectory (e.g., total interest paid, number of completed cycles). These reflect cumulative history and are useful for monitoring and early-warning systems but should not be treated as levers; attempting to change them directly confuses the symptom with the underlying cause.

Table 4 provides examples of these three categories.

TABLE 4. Examples of Actionable Levers, Signals, and Indicators

Source: Authors.

Before acting on any PCM finding, institutions should ask which of these three categories the finding belongs. The practical consequence of conflating these categories is significant and takes two forms. First, treating a milestone indicator as though it were a lever leads to investment in interventions that track success without creating it. The institution is following the signal rather than driving the outcome. Second, treating a signal as a lever leads to trying to change things the institution does not control. Signals belong in intake and targeting decisions, not in product design.

The critical distinction for assessment signals is that they should inform how clients are assessed and targeted. After developing a strong understanding of its PCM findings, an institution can use them—especially those related to actionable levers—to support institutional action at three levels. At the most immediate level, they can guide operational improvement now: directing staff attention toward the experience patterns associated with better outcomes, for example, adjusting credit ceilings and revising delinquency protocols (see Box 6). At the second level, the findings can target qualitative inquiry: launching a focused investigation to uncover the mechanisms and implementation differences behind the patterns the data reveal (see Box 7). At the third level, they can inform rigorous testing: codifying promising practice bundles identified by the analysis and testing them through controlled pilots or experiments. These three levels—improve now, explore why, test and scale—represent a progression from immediate operational adjustment through deepened understanding to validated, scalable change.

3.4 A Practical Note on PCM's Constraints

As with all methodologies, PCM has several constraints that FSPs and their funders and investors should understand before undertaking a PCM analysis.

First, as noted in Section 2.2, PCM findings sit along a spectrum of causal confidence. The strength of any given finding depends on data availability, coverage, and quality, as well as on how closely matched groups resemble one another. Where comparability is strong, causal claims carry greater credibility. While weaker comparability still yields useful associational evidence, that evidence requires more contextual interpretation before institutions act on it.

The methodology requires high-quality, longitudinal, client-level data that connects baseline characteristics to product exposure and outcomes such as client welfare.

Second, PCM hinges on data availability and quality. The methodology requires high-quality, longitudinal, client-level data that connects baseline characteristics to product exposure and outcomes such as client welfare. CGAP's pilots found that what appears to be a comprehensive institutional data set can prove fragmented in practice, with gaps in historical records, incomplete demographic data, and limited outcome data. Data fragmentation of this kind can affect the level of effort required; in CGAP's pilots, data preparation alone took one-third to one-half of the total analytical effort.

Third, PCM cannot explain why the patterns it finds exist. The analysis can reveal that a particular combination of product features (such as loan size or repayment structure) and accompanying services (such as financial education or peer support) is associated with better outcomes for a given client segment, but it does not provide insight into why these associations hold. For example, in one of CGAP's pilots, PCM found that center effectiveness was an important predictor of successful client outcomes among certain client groups. However, with the data available, center effectiveness was a black box; the analysis could see that it mattered, but not what specifically made a center effective. In such cases, PCM is directional: it can identify where to look and which questions qualitative follow-up should address.

Fourth, analytical findings gain institutional value only when they are translated into decisions that loan officers, branch managers, and product teams can act on. This translation from segment-level statistical patterns to operational guidance is distinct from the analytical work itself and typically requires deliberate investment. A first PCM analysis commonly identifies where operational refinement is most warranted, but developing the tools and processes to act on those findings is a further step. Communicating that nuance is itself a challenge, since findings differ by client segment, country, and product type. Translating that complexity into clear guidance that loan officers and branch managers can act on requires deliberate effort and investment.

Relatedly, while the value of PCM lies in its nuance—findings differ by segment, country, and product type—presenting segment-level point estimates without acknowledging their uncertainty can create false confidence. Analytical partners and internal teams should agree upfront on how to signal strength of evidence when presenting findings: at a minimum, this means distinguishing between findings based on well-balanced matched groups and those where comparability is weaker, and treating percentage-point differences as indicative of direction and approximate magnitude rather than as precise predictions.

3.5 Evaluative and Operational Applications

In its evaluative application, PCM generates evidence that informs periodic decisions about product design, pricing, targeting, and market entry. The analysis is conducted periodically and produces findings that leadership reviews and translates into policy. This evaluative application does not require integration with operational systems. It is the appropriate starting point for most FSPs and the mode in which CGAP's pilots operated.

In its operational application, PCM outputs are integrated into real-time or near-real-time systems supporting individual client decisions. This requires more sophisticated data infrastructure, real-time analytical pipelines, and careful attention to governance and ethical oversight. Operational applications offer the potential for much greater impact at scale, but they are considerably more complex and require either substantial experience with the evaluative application or external support for such work.

Ultimately, as noted in Section 1.5, understanding how to best utilize PCM is important for both applications. PCM is a diagnostic tool that identifies service delivery gaps based on outcome achievement with varying degrees of causal confidence (from strong to indicative). The findings it generates support informed decision-making by identifying where to look and what questions to ask. For actionable levers, they point toward changes worth piloting. For assessment signals and milestone indicators, they can inform better targeting of existing products, monitoring, and segmentation. In all cases, specific magnitudes of causal estimates should be treated as directional rather than guaranteed.

4. Is Your Institution Ready for PCM? A Practical Self-Assessment

A key question that many institutions may have upon learning about PCM is whether they have what it takes to make it work. The five conditions described below address that question. They should be read as a diagnostic, not a checklist. The purpose is to understand where the binding constraints are and what would address them. Most institutions will not satisfy all conditions fully at the outset; that is expected.

A word on resource expectations is also necessary. Experience from CGAP's pilots do not provide a comprehensive cost estimate. Institutional contexts and data environments varied too much for a single figure to be meaningful. In CGAP's pilots, data preparation consumed between one-third to one-half of the total analytical effort before modeling could begin. Institutions and investors should not budget for PCM as they would for a standard evaluation. The scoping phase alone requires significant time from senior analytical and domain experts. Accordingly, PCM should be viewed as an investment requiring both external and internal resources: methodological support from external technical partners and staff time to prepare data, engage with PCM analysts to guide key decisions, and help interpret findings. For more information, Annex D builds on this chapter by providing detailed data specifications, minimum quantitative thresholds, and pilot lessons for analysts and data managers preparing to undertake a PCM analysis for the first time.

4.1 The Five Conditions

1. Data volume and quality. PCM requires high-quality, longitudinal, client-level data, that is, records that track the same clients across time rather than capturing a single snapshot. The minimum viable data set links client demographics at intake, product and repayment history, and at least one outcome variable. In practice, many institutions find their MIS records only the most recent loan rather than the full loan history, or that demographic data was collected at intake but never updated. These are solvable problems, but they need to be identified before the analysis begins.

2. Meaningful outcome measurement. The target outcome should represent an important change in a client's situation, not an operational metric that describes product usage (such as repayment completion or loan renewal). Where an ideal target outcome cannot be directly measured, proxies can be used, but the distance between the proxy and the ideal target outcome must be acknowledged explicitly. FSPs that collect standardized data on income or business performance at each loan renewal may already have outcome panel data sufficient for a first PCM analysis.

3. Sufficient treatment variation. PCM evaluates variation in product experience across clients who all use the institution's services. If every client receives exactly the same product under exactly the same conditions, there is nothing to compare. In practice, most institutions have considerably more variation than they initially appreciate: product lines evolve, rollouts occur at different rates across branches, eligibility criteria create thresholds, etc. The scoping phase should include a systematic audit of available sources of variation and the data that captures it.

4. Analytical capacity. PCM requires statistical expertise that most FSPs do not yet hold in-house (see Section 1.7). This is not a barrier to getting started. CGAP's pilots were all conducted with external technical partners, and that model is well suited for a first use of PCM. The more important prerequisite is that internal staff understand PCM well enough to participate meaningfully in the interpretation phase; findings that only an external technical partner understands will not change operational practice.

5. Institutional change management capability. PCM findings will sometimes challenge established product designs and operational practices. An institution whose leadership approaches evaluation primarily as a compliance exercise will extract limited value. Institutional change management capacity means leadership that is willing to act on disaggregated evidence, that treats the question "which clients are we not serving well?" as a significant institutional priority, and that has processes in place to test product innovations informed by PCM findings.

4.2 For DFIs and Other Investors

For DFIs and other investors considering commissioning or funding PCM across a portfolio, the readiness assessment has two dimensions: the readiness of the specific FSP, assessed against the five conditions above, and the investor's own readiness to act on what PCM produces.

When assessing an FSP at the scoping stage, data quality and institutional change management capability are the two conditions from the five listed above that most reliably predict whether a PCM analysis can be conducted successfully and translated into institutional action. An institution with clean, longitudinal administrative data and leadership committed to acting on disaggregated findings is a strong candidate, even where analytical capacity is limited, as this gap can be addressed through a technical partnership. An institution with fragmented data or whose leadership approaches evaluation primarily as a compliance exercise is a weaker candidate regardless of data volume.

On the investor side, the material questions are whether the investor can engage with disaggregated findings rather than summary metrics; whether it is prepared to act on evidence—drawn from its own portfolio—that product design or delivery requires adjustment; whether its operating model can deviate from a one-size-fits-all model; and whether its reporting requirements and funding cycles can accommodate the multi-month timeline that a rigorous PCM analysis requires. Investors are best placed to benefit from PCM when the analysis is seen as an investment in institutional learning and grantee capacity building; they are interested in fund- and portfolio-level performance with greater precision than headline repayment and outreach metrics allow; and they seek a better understanding of how grantees can enhance product and service delivery for different client segments.

4.3 Indicators of a Productive First Use

A productive first PCM analysis should generate four outcomes: a clear mapping of client segments and the product configurations associated with positive outcomes within each segment; a quantified estimate of the delivery gap for each segment; a set of specific, operationally grounded hypotheses about what is preventing broader delivery of effective combinations; and at least one concrete institutional decision, such as a policy adjustment, a product modification, or a revision to data collection, that can be attributed to the findings.

4.4 What to Do Now If Your Institution Is Not Yet Ready

Many institutions will not be ready to undertake a PCM analysis immediately. The following data practices, informed by experience from CGAP's PCM pilots, build the foundation for future analysis and carry independent value for institutional decision-making regardless of whether PCM is subsequently pursued.

• Collect standardized information at each loan renewal, rather than at intake alone. Repeat assessments of income or business performance create a form of administrative panel data. Where possible, build in data collection of transformative outcomes that the organization would like to assess its ability to affect.
• Document the rules and practices that determine which product features clients receive. Data on treatment variation is essential for causal identification and is frequently not stored in MIS systems. For example, such information can include written credit policy manuals specifying minimum and maximum loan sizes by cycle, and any formula linking loan size to prior repayment history, collateral, or group performance; documented progression schedules²⁵ and any branch- or region-specific variants of that schedule; and product eligibility criteria that route clients into individual vs. group lending, or into products with different progression speeds.
• Begin systematically collecting data on elements of service delivery (e.g., client experience and satisfaction data such as staff competence, service speed, and fair treatment). Where this data was available in the CGAP pilots, these elements were highly predictive of successful outcomes.
• Track loan officer tenure and client assignment history. This data does not exist in most MIS systems and must be purposefully collected. Findings from one of CGAP's partner institutions suggest these variables may provide important equity signals.
• Leverage the emerging insights from CGAP's pilots to track specific elements that arose as predictive of successful outcomes:
  • Track delinquency trajectories, not just events. While most microfinance institutions (MFIs) monitor portfolio-level delinquency rates, fewer institutions systematically record client-level recovery patterns. Record whether delinquent clients are recovering or deteriorating, not just whether a payment was missed.
  • Record progression velocity explicitly, not just the current loan cycle or product tier. In CGAP's pilots, progression velocity was defined as the rate at which a client's loan size (or capital access) increased over successive cycles, relative to the rate observed among comparable peers with similar baseline risk and capacity profiles. How quickly a client is advancing relative to comparable peers is itself a meaningful predictor of outcomes, but most MFIs leave it as a derivable rather than tracked variable.

5. From Evidence to Action

Analytical findings do not improve client outcomes. Decisions do.

Analytical findings do not improve client outcomes. Decisions do. The most common failure in institutional data work is not methodological, but the gap between a well-executed analysis and an institution that does not act on the results. This chapter focuses on how to close that gap.

Closing this gap is meaningful for FSPs. Better visibility into which clients are and are not receiving product and service combinations associated with positive outcomes can translate directly into measurable improvements in portfolio performance metrics, such as lower non-performing loan rates (by identifying and addressing delinquency trajectories rather than single missed payments), improved client retention (by ensuring clients receive configurations matched to their circumstances), and stronger portfolio performance over time. Equally important, PCM also strengthens the broader credibility of the institution's model—commercial and mission-driven alike—and enables institutions to demonstrate not just that their products reach clients but that they deliver for them.

5.1 For FSPs

Start with questions, not methods. The most important preliminary step is identifying the analytical question the institution wants to answer. Productive PCM questions are specific: not whether a product works in general, but for which client segments particular product and service combinations are associated with better outcomes within the institution's portfolio. The question then extends to what the delivery gap is between those combinations and what clients are currently receiving. (See Annex A for more guidance on leading with questions, not methods.)

Delivery consistency before product redesign. The most consistent pattern across all five pilots was not that products were fundamentally wrong, but that their configuration (e.g., loan sizing, progression pace, delinquency tolerance) did not match what the analysis identified as associated with positive outcomes. Some of these product configuration constraints are product policy decisions: credit ceilings, standardized progression structures, uniform delinquency protocols. Others are operational: loan officer assessment quality, caseload pressures, incentive structures. Both types are largely within institutional control. Before investing in new product development, institutions should investigate which type of constraint is sustaining the delivery gap for that segment. The answer shapes where to direct the improvement effort.

Measure trajectory, not snapshots. In the cross-cutting findings on delinquency, loan sizing, and growth velocity that emerged from CGAP's pilots, the trajectory of a client's experience was consistently more predictive of outcomes than any point-in-time measure. Institutions that track whether clients are progressing or stagnating will generate better operational intelligence than those relying on snapshot metrics. This does not require new systems; it requires structuring existing data to capture movement over time.

Reorient analytics toward client outcomes. Many financial institutions possess substantial analytical capacity, but that capacity is typically oriented toward credit risk management—asking which clients are likely to default. PCM complements this perspective by asking which clients are likely to experience the welfare improvements the institution's products are designed to produce. These are related but not identical questions. For many institutions, this means redirecting existing analytical capacity toward these questions, augmenting existing capabilities rather than building new capabilities from scratch.

Address known implementation barriers. Three barriers arose consistently across the CGAP pilots. The first is communication complexity: PCM surfaces detailed findings about multiple client segments. The sheer amount of information can be overwhelming for an institution accustomed to metrics that are easier to understand at first glance. The second is integration into existing systems: findings that exist in a report but cannot be accessed through the tools staff use in daily practice will not change operational behavior; segment-specific findings must be translated into decision rules, reference cards, or flags within existing systems that loan officers can use under time pressure. The third is resource constraints: institutions should distinguish between changes that require significant resource commitment, such as creating new programs, and those that principally require a policy adjustment decision, such as revising credit ceilings, changing product availability by region, and adjusting delinquency response protocols. Institutions that make this distinction early can act on low-cost policy changes immediately, while building the case for the larger investments over time.

5.2 For DFIs and Other Investors

Frame PCM as an investment for institutional learning and internal capacity building, not a funder or investor evaluation. An investor-funded PCM analysis is best framed as an investment in the FSP's own analytical capacity and ability to improve successful outcomes for clients, not as an external evaluation conducted for funder accountability. Evidence designed to satisfy funders' needs rather than the questions FSPs are seeking to answer is less likely to shape operational decisions. Instead, PCM analyses designed around the FSP's own analytical questions, conducted with its staff as active participants, and concluded with a learning session that connects findings to specific institutional decisions are substantially more likely to generate durable change than those conducted at arm's length and reported upward. DFIs and other investors can embed these design principles into their technical assistance grants and scoping conversations.

Be realistic about where value is initially realized. How far PCM findings travel up the investment chain remains an open question. Impact-oriented investors²⁶ should be cognizant that the value of PCM will initially be realized at the FSP level, not the fund level. Investors can support their investees by building PCM into multi-year technical assistance programs and designing reporting requirements that create incentives for delivery gap reduction alongside average impact improvement.

5.3 Using Findings Well: Common Errors to Avoid

Confusing predictive models with approaches like PCM is perhaps the most consequential analytical error. A predictive model identifies characteristics that correlate with client success but does not identify what caused that success. Such models show how outcomes are currently distributed. PCM integrates a causal framework into the modeling process to produce insights that identify where effects differ across client segments and generating grounded hypotheses about causality.

Overinterpreting noisy estimates is a common error in machine learning applications. Random forest estimates for small client segments carry wide confidence intervals. Institutions should be attentive to the uncertainty around estimates, not only their point values. This is reflected in the guidance to use PCM's findings as structured, directional hypotheses.

Ignoring the limits of the causal identification strategy undermines the causal claims that distinguish PCM from simpler approaches. Institutions should require technical partners to be explicit about the strength of causal credibility, based on how similar matched groups are, and should calibrate their own confidence in findings accordingly.

Treating PCM as a one-time exercise misses its principal value. A single use produces a snapshot that, however well executed, generates limited institutional value unless it contributes to an ongoing analytical learning cycle. The value compounds when PCM is repeated: each cycle validates, extends, or revises the previous findings, building a progressively richer understanding of the institution's client population.

Conflating program-evaluation findings with operational decision rules is an error with practical consequences. PCM at the program-evaluation stage identifies which product features and service approaches are associated with better outcomes for clients in comparable circumstances. It does not specify what to do for a particular client applying for a specific loan on a given day. The translation from segment-level evidence to individual decision guidance requires a deliberate additional step involving operational staff.

5.4 Building Sustained Analytical Capacity

As noted, while the evaluative application of PCM is a valuable starting point and can provide useful insights for an institution, the goal is to move from periodic commissioned analyses to an embedded practice in which evidence from institutional data regularly informs product decisions. FINCA International tested elements of in-house analytical replication during its pilot, providing early evidence that this transition is achievable with the right resources and commitment (see Box 4). The transition requires data systems configured to preserve longitudinal records, clearly defined analytical roles, revised decision-making processes, and a culture that normalizes evidence-based adaptation.

Conclusion: The Case for Precision

For FSPs, PCM's central lesson is that the information needed to improve client outcomes—to match product design and delivery to what works for each client segment—is, in many cases, already embedded in an institution's own data. PCM can make these insights visible and connect them to the product, policy, and operational decisions within institutions' control. This is a meaningful departure from how the sector has typically generated and used evidence.

Over the past two decades, the financial inclusion sector has invested substantially in generating evidence, yet its practical utility for FSP decision-making has remained limited. Average impact findings tell institutions that their products were broadly effective for the population studied. PCM offers a different kind of evidence: not average effects, but the pattern of effects across different types of clients and circumstances; not whether the product works, but which configurations of product features work and for whom. It is equally demanding to produce as other methods but can be more directly useful for the decisions facing FSPs and their investors.

The CGAP pilots gave concrete form to PCM's potential. In every institutional context studied, the majority of clients faced a delivery gap: They were not receiving the combination of services the data identified as most strongly associated with positive outcomes for people in their circumstances. The scale of that gap and its operational origins make the case for action. The five pilots revealed an emerging set of cross-cutting insights that can serve as a guide: Under-lending was consistently associated with poorer outcomes than appropriately sized loans; delinquency trajectory mattered more than occurrence; growth velocity and progressive structuring were associated with success; and how services were delivered mattered, not just what was delivered. These findings are not prescriptive. They are patterns, observed across enough contexts, with enough specificity, to be useful as a structured set of questions to ask of an institution's own data.

CGAP began this work looking for an evaluative methodology that would enable institutions to understand what works for whom in a given portfolio. What the pilots revealed was something with greater practical potential: a methodology that operates less as a retrospective verdict and more as an ongoing decision-support process, enabling tighter feedback loops between analysis and action. Such a shift from evaluation to decision-support offers a more direct path from evidence to impact. PCM is just one example of what is becoming possible as machine learning, theory-driven program design, and strengthening data infrastructure converge. The sector's task is not to standardize on any single method, but to build the analytical culture and data foundations that allow methods like PCM to fulfill their potential.

Annex B – The Six Phases of PCM

A PCM analysis proceeds through six phases. This annex describes these steps in greater detail and is intended primarily for analysts and technical staff who will design or execute a PCM analysis, as well as institutional leaders who want a detailed understanding of what each phase involves. Figure 1 (page 11) provides a high-level overview of the PCM process, from inputs to analytical modeling to interpretation.

Phase 1: Theory of Change and Target Outcome

Before any data are assembled or any model is specified, the parties to the PCM analysis need to agree on two things: what the financial services in question are trying to achieve (the "target outcome") and how they are supposed to achieve it (the theory of change). The target outcome should represent meaningful change in a client's situation. Repayment completion, loan renewal, or product take-up are operational metrics; they tell the institution about client usage, not whether the product helped. The question to ask is: what does a client's life or livelihood look like when the product has truly worked?

Several tensions arise in practice: 1) the time horizon problem, as the outcomes that matter most take longer to emerge than data windows allow; 2) the gap between what the FSP wants to know and what data allow; and 3) institutional tensions where different stakeholders have different views about what success looks like. PCM forces these disagreements into the open at the start of the process.

The target outcome selection is critical because it influences which intermediate outcomes emerge as predictive of achieving the target outcome in the modeling, anchors the ideal outcome modeling step (described in Phase 3), calibrates how findings should be interpreted, and, where the gap between target and available data is large, generates a concrete investment case for the institution's data agenda (see Box B1).

The theory of change does two distinct jobs in PCM: one before the modeling begins and one during it. Before any data are assembled, agreeing on the theory of change forces the analysts and FSP staff working together on the analysis to specify the causal pathway through which the product is expected to produce the target outcome. This is important, as it surfaces institutional disagreements about what the product is for, which stakeholders often discover they hold different views about only when forced to make them explicit. Getting this wrong at the start, or skipping it, means the subsequent modeling lacks a principled structure.

During the modeling, the theory of change plays a technically important role: it determines how variables are tagged. Every variable in the data set is categorized into one of five types: context variables (pre-existing conditions outside the FSP's control), baseline variables (client characteristics at entry), experience variables (the specific product and service elements the client received), outcome variables (intermediate results on the pathway to the target), and the target outcome. This tagging is what prevents the random forest modeling from scanning blindly across all available variables and confusing causes with effects. It preserves the logical sequence of the causal pathway so program activities are not mistaken for their consequences and confounding pathways are blocked before modeling begins.

The practical implication is that an institution that cannot articulate a plausible causal pathway from its product to its intended outcome will struggle to tag variables correctly, which compromises the integrity of the modeling.

Box B1. Lessons from the CGAP Pilots: Target-Setting

The choice of target outcome proved to be the most consequential decision in each of CGAP's PCM pilots and consistently one of the most difficult. Behavior-based targets such as FINCA's financial wellbeing composite (which uses indicators related to three key dimensions: 1) managing needs; 2) coping with shocks; and 3) building for the future) and the Banco Central do Brasil's financial resilience proxy (constructed by CGAP from behavioral fast payments usage) tended to surface operational variables at the top of the analysis—the use of savings products, product experience quality, service patterns—that can map more directly to institutional decisions. Poverty-based targets, used for AMK Cambodia and Grameen Costa Rica, surfaced more contextual and geographic variables. But this is a tendency demonstrated by the CGAP pilots, not a rule, and can be influenced by the data available for analysis and the design approach (episode- vs. window-based): Grameen's analysis, which drew entirely from administrative data, produced highly actionable findings with a poverty-based target, and a second modeling of AMK's data using its pre-existing client personas showed that the segmentation framework matters as much as the target. Each target type brought trade-offs in what it measured and what it surfaced, but all five produced meaningful findings with substantial variation across client segments. The most durable lesson: start with what transformation looks like for the client, then work backward to measurement. Do not anchor to what is already tracked if it does not reflect transformative change.

Phase 2: Data Assembly and Quality Assessment

The core data requirement is a longitudinal client-level data set linking demographic characteristics at entry, product uptake history, and outcomes over time (see Box B2). Assembling these data typically requires extracting records from multiple systems, resolving inconsistencies, constructing consistent variable definitions across time, and assessing quality and coverage. Data quality assessment should address four dimensions: completeness, accuracy, longitudinal consistency, and coverage. This assessment does more than establish analytical limits; it generates a prioritized agenda for improving data systems.

Treatment variation data (i.e., records of which specific product features different clients actually received) should be considered at the scoping stage. It is frequently not stored in MIS systems and has to be assembled from operational records, product documentation, or internal policy archives. Without it, PCM cannot identify which configurations are associated with better outcomes because it cannot see that configurations varied. In practice, treatment variation is almost never truly absent; even in one of CGAP's pilots, in which there was a standardized lending model, there was variation in progression velocity, timing, and discretionary overrides, which can serve as critical treatment variation data in PCM modeling. As such, at the scoping stage, institutions should explore whether variation exists, confirm that the data is retrievable (from operational records, etc., when the MIS does not track it cleanly), and confirm there is sufficient spread across likely comparison groups to support within-group testing. In summary, institutions that discover this gap during scoping should not treat it as disqualifying; it is solvable, but it needs to be identified early.

Common problems across the CGAP pilots included MIS systems recording only the most recent loan rather than full loan history, demographic data collected at intake but never updated, outcome variables recorded inconsistently across loan cycles or branches, and insufficient documentation of eligibility rules and product assignment mechanisms. Records that appeared rich through a reporting interface regularly proved fragmentary when assembled for analysis. Data preparation consumed between a third and a half of total analytical effort in every pilot. It is not merely a preliminary step before the real work begins, but core to the work itself.

Box B2. Lessons from the CGAP Pilots: Surveys Are Not Required—But Longitudinal Data Is

Asociación Grameen Costa Rica came into the project believing the absence of surveys was a significant disadvantage. What the research team found was that their clean administrative data was sufficient. What PCM requires is longitudinal data with enough depth to observe patterns of behavior across a client's relationship with the institution. A single loan with rich data collected at known intervals could in principle provide sufficient depth. What does not work is a single snapshot with no trajectory information. This finding substantially lowers the feasibility barrier for applying PCM.

Phase 3: Creation of Block Matched Groups (Comparable Client Groups)

A critical point of orientation: there is often no "control" group in the data set. Every record belongs to a client of the institution. What PCM evaluates is variation in product experience (e.g., differences in loan terms, repayment structures, product combinations, or service features) across clients who are all using the institution's services.

Before constructing comparison groups, PCM first models how context, baseline, and experience relate to the target at the population level. Flexible tree-based models are used to estimate, for each client, a general response score: a summary of how strongly that client appears to be moving toward the target outcome, given their observed experience profile. Parallel models capture how context and baseline relate both to this general response and directly to the target outcome, providing a compact description of each client's developmental response architecture without yet defining any "treatment."

Using these population level measures as inputs, PCM then constructs BMGs: sets of clients who start from broadly similar contextual and baseline conditions and exhibit similar overall response profiles before any specific difference in product experience is considered (see Figure B1 for an example of a BMG decision tree and Box B3 for insights from CGAP's pilots). Within each BMG, clients are therefore comparable on the factors most likely to confound experience–outcome relationships and remaining systematic differences in exposure to key product features can be examined more credibly. The approach still rests on a selection on observables assumption—that the characteristics systematically influencing which clients received which features are captured in the data—but it makes that assumption explicit and testable through balance diagnostics and overlap checks at the group level.²⁷

Figure B1 – BMG Decision Tree (Illustrative)

Source: Authors.

A note on demographic variables. A deliberate design decision arises at this stage: whether to include demographic characteristics—such as gender, ethnicity, and geographic location—in the block match model or to exclude them. Each choice serves a different analytical purpose. If demographic variables are excluded from block matching, they remain available as independent analytical lenses in the equity analysis phase (Phase 6). This sequencing allows the analyst to ask, after identifying which product configurations are associated with better outcomes, whether those configurations are equitably distributed across demographic groups and, separately, whether those who do receive them achieve comparable outcomes.²⁸ This is the approach taken in CGAP's pilots (in which gender variables were withheld for most partners²⁹) and is recommended for any institution, commercial- or social-mission-driven, where equity of access is a core investigative question.

Alternatively, if demographic variables are included in block matching, they are treated as potential drivers of outcome heterogeneity on equal terms with other baseline characteristics. This approach is appropriate where the institution wants to test whether, for example, gender or location is an important moderator of which product configurations work, rather than treating it as an equity lens applied after the fact.

The choice between approaches depends on what answers the institution wants to surface: if it wants to examine whether specific demographic groups receive the product configuration that is most effective for their BMG and whether they achieve comparable outcomes to others within the BMG, the demographic variables in question should be removed from the modeling in Phase 3; or if the institution wants to assess if that demographic variable is a driver of which product configurations work, then the variable should be included in the modeling. Critically, the two approaches are not mutually exclusive: analysts can run parallel models, one including and one excluding demographic variables, to understand both what demographic groups predict about outcomes and how equitable the distribution of effective configurations is across them. This requires two different sets of models, but once Phases 1 and 2 are complete, this can be a manageable lift. Whichever approach is chosen, it should be decided explicitly at the design stage and documented alongside the rationale, since it affects how Phase 6 findings can be interpreted.

Box B3. Lessons from the CGAP Pilots: Existing Segmentation Frameworks Can Be as Rigorous as Data-Derived Ones

In modeling with AMK Cambodia's data, the team found that existing institutional categories (client personas) can be as powerful as data-derived segments. The initial analysis used data-derived segments that produced seven groups with dramatically different success rates, but AMK could not act on the findings because the segmentation variables, drawn from a national database were not in their operational data. A second model using AMK's existing personas was analytically comparable and operationally superior. The test for any segmentation approach is whether the staff making decisions about a client can readily identify which segment that client belongs to.

Phase 4: Estimating Heterogeneous Treatment Effects

Within each BMG, PCM focuses on experience combinations rather than individual variables in isolation. Flexible tree-based models (specifically random forests) are used to learn how experience variables relate to the target for clients within that group, and to identify the experience pattern that is most strongly associated with higher target outcome achievement among comparable clients. The random forest builds many decision trees in parallel, each trained on a random subsample of the data and a random subset of features, then averages across them. This approach balances pattern-detection flexibility, bias reduction, and scalability, and is particularly well suited to detecting interactions between variables that simpler models would miss: for instance, that larger loans were associated with success for established business owner clients (one BMG) for one of CGAP's partners but not for its casual-labor clients (another BMG) when modeling was conducted by livelihood persona.

Each client in the BMG is then classified as either having received the identified pattern or not, and simple statistical tests (such as Fisher's exact test) compare target achievement rates between the two sub-groups. This produces a clean, interpretable contrast: within this group of comparable clients, those who received this configuration achieved the target outcome at X percent, compared to Y percent for those who did not. Results are aggregated across BMGs to produce population-level findings that preserve contextual comparability. SHAP values are used alongside this to attribute predictive importance to individual variables within each modeling step, identifying not just which overall experience pattern is associated with success but which specific elements within it did the most explanatory work.

Because the discovered pattern is a data-driven contrast rather than a pre-assigned treatment, PCM includes a separate population-level step to test whether receipt of the pattern is itself confounded by context and baseline characteristics: an out-of-fold propensity model for pattern receipt is fitted using context and baseline variables, overlap and weighted balance between recipients and non-recipients are checked, and a credibility label of strong, moderate, or weak is assigned. This step—not the random forest or the Fisher's exact test alone—is what determines how much causal weight a given finding can bear. This is what allows findings to be categorized into the levers, signals, and indicators framework in Phase 5 (see Section 3.3).

The distinction between episode-based and window-based approaches to PCM—along with the choice of target outcome—affects what Phase 4 tends to surface. In episode-based analyses from CGAP's pilots (three institutions), which focused on poverty reduction target outcomes, the experience variables span the entire client episode from first loan through to the completion point, capturing cumulative trajectory patterns such as how loan sizing evolved, how progression velocity developed, and how delinquency was handled across the full arc. This approach appears more suited to deep analysis of specific cohorts. In window-based analyses from CGAP's pilots (two institutions), which focused on financial health and financial resilience, the experience variables reflect what happens to clients during the defined window, and findings tend to foreground variables like savings product usage, service quality dimensions, and payment behavior patterns—things observable within a defined period that map naturally to ongoing operational decisions about how to manage existing client relationships.

Two important caveats apply across both approaches. First, the group-specific effects carry uncertainty. They indicate directions and approximate magnitudes of heterogeneity, not precise individual-level predictions. Second, every modeling step uses data partitioning and k-fold cross-validation to guard against overfitting and, specifically, out-of-fold predictions to prevent data leakage in the developmental-response measures: the data set is split into training and testing subsets; no case's predicted value is generated by a model trained on that same case, and performance is averaged across multiple folds to confirm that patterns generalize rather than reflect noise in a particular subsample. This is what distinguishes PCM's machine learning component from simpler predictive analytics.

Phase 5: Interpretation and Application

The translation from statistical findings to operational conclusions requires deep knowledge of the FSP's products, client base, operations, and strategic context. The most effective implementations involved iterative collaboration between technical analysts and operational staff throughout the analysis, not just at the interpretation stage. It is critical to establish actionability goals before the analysis begins: "What decision will this finding change, and who will make that decision?" should be part of the study design conversation, not just the debrief.

Phase 6: Equity Analysis

An equity analysis examines whether the benefits of the product feature being evaluated are distributed equitably across demographic groups within each BMG. Where demographic variables were excluded from the block-matched group model in Phase 3, as recommended when equity of access is a core investigative question, they can be used here as independent lenses. Where demographic variables were instead included in Phase 3, equity questions are addressed directly through the heterogeneity results, rather than through this separate step.

While not a mandatory component, equity analysis is highly recommended for any institution, commercial or social-mission-driven, for whom equitable outcomes are a priority, as it provides important information on where the institution can do more to improve outcomes across certain types of clients (for example, women). Discovering that a product feature generates strong effects overall while consistently failing to reach a specific client segment is directly relevant to both product design and targeting decisions.

1. PCM was developed and tested by BCT Partners, in consultation with government and private funders, to help address complex social issues through the use of large-scale data.
2. CGAP's broader examination of the impact literature related to credit can be reviewed on the Impact Pathfinder platform and in Opening the Black Box on the Impact of Inclusive Credit (Nielsen et al. 2026). The evidence synthesized in these publications examines the impact of inclusive credit. Much of the literature referenced in these two publications refers to microcredit, delivered by microfinance institutions, but some refers to credit more broadly. In all instances, the literature included focuses on credit products provided to individual borrowers, households, or micro and small enterprises in low- and middle-income countries. As such, CGAP uses the term "inclusive credit" where findings apply across product types.
3. The Impact Pathfinder can be found at https://www.impactpathfinder.org/.
4. The views and conclusions expressed in this publication are solely those of CGAP and do not necessarily reflect the views or positions of the Banco Central do Brasil (BCB). While BCB contributed with data to the analysis, such contribution should not be interpreted as an endorsement of the methodology, analysis, or conclusions presented.
5. RCTs using administrative data can be as fast and inexpensive as observational approaches, and machine learning methods for heterogeneous treatment effects can be applied to RCT data as readily as to observational data. The real cost driver, in general, is obtaining high-quality longitudinal outcome data, not the choice of method. The cost and heterogeneity limitations discussed here apply specifically to conventional survey-based RCT designs, which remain the modal approach in the financial inclusion sector.
6. These institutions were selected through a competitive expression of interest process designed to identify institutions with the data infrastructure and institutional commitment needed to test PCM's viability adequately. A key selection criterion was the availability of outcome data or proxy outcome data, not just operational records, which is rare among FSPs and essential for analysis of client-level outcomes.
7. This is done through tests examining selection-on-observables and sensitivity to unobserved confounders.
8. A random forest is a machine learning algorithm that builds and combines many decision trees to generate predictions (see Key Terms). PCM leverages random forests because the model balances causal fidelity, bias reduction, and scalability, all of which are critical for a framework in which missing data can derail confounder control, subgroup analysis, and the validity of subsequent causal claims.
9. SHAP (SHapley Additive exPlanations) is used to help identify the top predictors in certain PCM modeling steps. SHAP is an explainable AI method that leverages game theory to explain predictions from machine learning models (SHAP n.d.).
10. Section 4.1 and Annex D provide fuller accounts of the data and other requirements for PCM.
11. To ensure time-varying effects are considered in the analysis, the episode-based approach draws from a specified time period (typically two to five years). As such, while the timing for individual client journeys may not be exact, it is relatively close. In such cases, PCM analysts will collaborate with the financial institution to identify whether there were any major changes (internal or external) that could affect block matched group assignment.
12. FINCA's Mission Monitor is an annual survey that helps the organization track whether it is achieving its social mission. It monitors FINCA's outreach to financially excluded populations, their use of savings and credit, and their ability to manage day-to-day needs and longer-term goals (FINCA International 2019).
13. The target was a financial resilience proxy constructed by CGAP from behavioral fast payment usage.
14. Uncovering the mechanism behind this finding will require further qualitative inquiry from the financial institution. Information on how to interpret PCM findings is in Section 3.3, and more on PCM's complementary nature to qualitative inquiry is in Box 5.
15. As measured by metrics related to staff warmth and respectfulness, ability to explain products, speed of service delivery, and fairness of treatment.
16. Data availability varied by partner. Product feature variables available in most pilots included loan amount/size, interest rate, loan terms and repayment schedule, loan purpose, number of loans/loan cycle, repayment behavior, fees and charges, and loan balance/outstanding debt. Other variables were available for specific partners.
17. A key question that arose with this finding was why accumulated microfinance expertise becomes a constraint for clients ready for commercial-scale products. PCM cannot provide insights into the mechanisms underlying the results; it can only surface what the data show. Researchers and financial institution staff discussed hypotheses that could explain this finding, but ultimately this finding represents an opportunity for the financial institution to conduct further qualitative research to better understand what drives it.
18. One of CGAP's partner institutions operates a group lending model organized around "Centers" that hold regular meetings of approximately 15–30 borrowers facilitated by a loan officer. Centers serve as the primary point of service delivery, combining loan disbursement and repayment with peer support, financial education, and business mentoring.
19. Center effectiveness was not captured by a single fixed rating; it was measured relative to each client segment's specific needs. For each of the matched groups in this analysis, a regression tree was used to predict which Centers were most associated with better client outcomes, grouping Centers into up to eight tiers based on their predictive strength. Critically, the same Center can be highly effective for one segment but not another. For example, a Center might be very effective for established business owners, but not effective at all for individuals living in districts with universal internet connectivity and moderate infrastructure access. This showed that effectiveness is not about the Center itself in isolation; it is about how well a particular Center's environment, leadership, and peer dynamics match the needs of a specific client population.
20. These examples represent findings from across the five pilots and do not apply universally to all partner institutions.
21. This draws from Judea Pearl's structural causal models—a mathematical framework that represents causal relationships between variables as a graph identifying which variables causally influence others, allowing researchers to distinguish between mere correlation and true causation. Structural causal models enable analysts to predict what would happen if a variable were intervened upon (not just observed), and to identify which variables must be controlled for to isolate genuine causal effects.
22. This borrows selection-on-observables logic (from the Rubin–Imbens potential outcomes tradition) to assess the causal credibility of discovered patterns through out-of-fold propensity modeling, common-support diagnostics, and weighted balance checks, addressing certain matching-quality concerns raised by King and Nielsen (2019).
23. PCM limits concerns related to the multiple testing problem through a locked, sequential analytic workflow. PCM has a defined workflow: target construction, developmental-response profiling, comparison group formation, within-BMG discovery related to outcome achievement, and causal credibility assessment. Each part of the modeling includes checkpoints for correctness, feasibility, and fit with the client context, but those checkpoints occur before results are known and are not used to rework the analysis in search of stronger findings.
24. The research design that incorporates specific assumptions and is used to establish causal relationships from data, especially observational data.
25. For example, a schedule might read: "loan cycle 2 = 130% of cycle 1 if no delinquency."
26. CGAP research has found that many impact investors continue to rely on outputs and indirect measures, with disaggregated outcomes data not yet a mainstream part of portfolio reporting (Lahaye, Clarke, and Kiamba 2024). It is likely that the appetite for segment-level findings will be greater among investors already innovating on impact measurement. For such investors, PCM offers a way of better understanding how well a grantee's services are achieving the intended outcomes across different client segments and how to enhance service delivery for different types of clients.
27. Every supervised machine learning step in PCM uses two foundational practices to avoid challenges to the validity of results: data partitioning and k-fold cross-validation. First, data partitioning splits the data set into training and testing subsets, typically at a 70:30 ratio. The training set builds the model, while the testing set acts as an independent "reality check" to evaluate performance on unseen data. This separation prevents overfitting, in which a model memorizes noise or idiosyncrasies in the training data rather than learning generalizable patterns. Without this split, models might falsely inflate confidence in variables that coincidentally align with outcomes in one data set but fail elsewhere. Second, K-fold cross-validation further strengthens reliability by rotating which portions of the data serve as training and testing sets across multiple iterations (e.g., 5-10 "folds"). In each fold, a different subset is withheld for testing, while the remaining data trains the model. Performance metrics are averaged across all folds, revealing how consistently the model performs under varied conditions.
28. These are distinct questions. Inequitable access with equitable outcomes points to a delivery problem; equitable access with inequitable outcomes points to a design problem in which the configuration may not work equally well across groups. In one CGAP pilot, for example, PCM revealed significant variation in gender equity across a single institution's countries of operation, identifying not only where gaps existed but what might be driving them, such as fewer women receiving the product and service combination most associated with positive outcomes. This level of precision carries risks as well as opportunities. More granular segmentation can help resolve both design and delivery problems, but it can also make it easier to exclude clients perceived as less likely to succeed, particularly when institutional incentives or biases are not well aligned with inclusion goals.
29. Asociación Grameen Costa Rica's services are only available to women, so no gender data was included in this analysis.
30. These dimensions were "managing needs", "coping with shocks", and "building for the future."