Executive summary

Insurers are significantly increasing their allocations to private credit, yet many privately placed structured-finance transactions lack the modeling files, standardized disclosures, or performance histories typically available for public deals. To address these challenges, we developed a nearest-neighbor (NN) modeling framework that systematically matches each unmodeled private deal to the most comparable public deal using five economically meaningful characteristics.

The framework provides a transparent, defensible, and scalable method for projecting cash flows without bespoke modeling.

Key takeaways

• NN mapping offers a scalable solution for unmodeled private deals: Our methodology identifies the most representative public deal using transparent, economically grounded matching criteria.
• Five characteristics ensure structural and economic comparability: Vintage, weighted-average coupon (WAC), credit quality, weighted-average life (WAL), and collateral type jointly form a standardized “deal fingerprint.”
• Strong validation supports reliability: Proxy cash flows differed from true modeled results by only 2.74% in present value, demonstrating high fidelity.
• More efficient than bespoke modeling: This approach provides credible analytics without the time, cost, or complexity of constructing individual models for each private deal.

Introduction: Why a nearest-neighbor approach helps insurers evaluate private credit deals

Private credit has become an increasingly important component of insurers’ investment strategies, driven by attractive yields, structural protections, and diversification benefits. However, many privately placed structured-finance transactions lack the complete information provided in SEC-registered public deals, such as waterfall files, collateral performance histories, or agency-modeled cash flows. These gaps introduce challenges for valuation, capital modeling, asset-liability management (ALM), and enterprise risk analysis.

Traditional approaches—building custom models for each private deal, using simplified heuristics, or limiting modeling altogether—either impose excessive cost or compromise analytical rigor. To help insurers bridge this gap, we have developed a nearest-neighbor (NN) approach that leverages public-deal data to infer credible cash-flow projections and risk metrics for private transactions.

Market context: Why private-credit modeling gaps exist

The global private-credit market has expanded rapidly, transforming the landscape of structured-finance investments. Private credit assets under management (AUM) increased from roughly USD $260 billion in 2010 to nearly $1.6 trillion in 2022. Private securitizations, including 144A placements, club transactions, and customized structures, represent a growing share of total issuance volume. Insurance companies now hold over USD $220 billion in privately placed asset-backed securities (ABS), illustrating the scale of the modeling challenge.

Yet private transactions often provide:

• Limited or nonstandard waterfall documentation
• Incomplete or inconsistent loan-level reporting
• Limited historical performance data
• No agency-modeled cash-flow files

As a result, insurers need a practical, standardized method for integrating private deals into valuation frameworks and risk models without sacrificing analytical depth.

Methodological overview: Helping insurers understand the nearest neighbor framework

Our NN framework identifies the most representative public deal for each private deal by comparing them across five core characteristics. These characteristics capture the structural, credit, and cash-flow attributes that most influence performance and were selected to align with insurer valuation needs, risk-based capital (RBC) frameworks, and observed structured-finance behavior.

Vintage

Vintage reflects the macroeconomic and credit environment at origination and influences collateral quality, underwriting standards, and borrower performance. Deals issued during stressed periods often exhibit different delinquency and loss dynamics than those issued during expansionary markets. By aligning vintages, we ensure matched deals share similar economic backdrops and expected performance patterns.

Weighted-average coupon (WAC)

WAC is a key driver of prepayment and refinance behavior. Higher-coupon pools tend to refinance more quickly in declining-rate environments, while lower-coupon pools amortize more gradually. Matching on WAC ensures both private and public deals exhibit similar rate sensitivity and cash-flow timing across a range of interest-rate scenarios.

Credit quality

Credit quality, expressed via numeric rating equivalents, reflects the expected loss profile of the collateral. It influences credit enhancement, expected default behavior, and RBC treatment. Incorporating credit quality into the match ensures that proxy deals reflect comparable levels of risk under both baseline and stressed analyses.

Weighted-average life (WAL)

WAL measures the expected timing of principal repayment and provides insight into extension risk and sensitivity to late-cycle macroeconomic conditions. Shorter-WAL structures return principal more quickly, while longer-WAL deals retain exposure to multiple economic cycles. Aligning WAL ensures accurate timing of projected cash flows and supports consistent ALM treatment.

Collateral type

Collateral type determines the structural mechanics—waterfall rules, loss allocation, prepayment patterns, and credit enhancement—underpinning the deal. Residential mortgage-backed securities (RMBS), commercial mortgage-backed securities (CMBS), ABS, and collateralized loan obligations (CLOs) exhibit fundamentally different behaviors. Exact matching on collateral type preserves structural comparability and ensures the model applies appropriate performance assumptions.

Matching process and proxy selection for insurers’ private credit deals

Each private deal is transformed into a parameter vector encompassing the five matching characteristics. Distances between the private and public deal universe are computed using a scaled Euclidean metric to balance the contribution of each variable. The public deal with the smallest distance is selected as the proxy, subject to exact matching on collateral type.

We considered matching exactly on both collateral type and rating to improve match quality, but doing so may lead to sparse data within each bin, causing many asset deals to be discarded.

To fine-tune the matching process, one can:

• Give more importance to certain variables by applying weights. For example, if credit quality matters more than vintage, you simply assign it a higher weight, so the algorithm is more sensitive to that difference.
• Set calipers, which are maximum allowable gaps between two deals on a given variable. If the distance exceeds the caliper, the pair is rejected.

These tweaks usually produce higher-fidelity matches, but they also can result in more securities that cannot be matched given the constraints. This can be addressed through gradual relaxation of constraints over multiple passes, at the expense of the match quality in each iteration.

Illustration of nearest-neighbor matches when evaluating private deals

We evaluate the quality of the matching by comparing the characteristics of the deal to those of the proxy deals generated through nearest-neighbor matching (NNM) and analyzing the distance measure between them to gain insight into the accuracy and reliability of the matching framework.

Figure 2 presents five illustrative matched pairs, highlighting the differences between the attributes of the deals and those of the proxies assigned by the NNM procedure.

For each matched pair, the table displays the scaled Euclidean distance metric alongside the characteristics of the deal and its assigned proxy, including rating, year, WAL, and WAC. Smaller distance values indicate a greater similarity between the actual deal and its proxy. Overall, the results demonstrate that the NNM procedure generally provides a close approximation to true deal characteristics.

Testing the nearest-neighbor match – validation and performance assessment

We validated the NN approach using a structured portfolio for which full modeled cash flows were available. Half of the portfolio was treated as unmodeled, matched to proxies, and valued using NN-generated cash flows. The present value of proxy cash flows differed from true modeled results by only 2.74%, demonstrating strong alignment with fully modeled analyses.

Cash-flow paths for proxy and true deals aligned closely across the projection horizon, capturing the timing, magnitude, and shape of principal and interest flows with high accuracy. These results support the suitability of the NN approach for valuation, capital modeling, ALM, and enterprise risk.

Conclusion: Nearest-neighbor modeling helps insurers expand into private credit

As private credit continues to grow across insurer portfolios, organizations require modeling approaches that balance accuracy, transparency, scalability, and cost-efficiency. The NN framework provides a reliable and analytically robust method for projecting cash flows and assessing risk for private structured-finance assets with limited disclosure. The approach enhances valuation consistency, supports model governance expectations, and enables reliable integration of private credit into capital, ALM, and enterprise risk models.

This methodology offers a practical and scalable foundation for navigating the evolving structured-finance landscape and addressing data limitations inherent in private-credit markets. Won’t you be my neighbor?

Appendix – A deep dive into the mathematics

Matching methods require the specification of a distance metric to identify the closest non-focal unit(s) for each focal unit or to serve as the criterion in an optimization procedure. This distance can be defined in terms of propensity score differences, which capture the relative position of units within the estimated probability space of treatment assignment, or through pairwise distances computed directly from the covariates. According to King and Nielsen,¹ pairwise distance methods tend to produce closer and more comparable matches than propensity score-based methods, as matched units are similar across all covariates, whereas units matched solely on the propensity score may differ substantially on individual covariates despite having similar scores.

Typical examples of pairwise distance metrics include Euclidean, scaled Euclidean, Mahalanobis, and robust Mahalanobis distances. These measures are computed for each treated unit i and a control unit j by

where x_i is a p×1 vector containing the value of each of the p included covariates for that unit and S^-1 is the (generalized) inverse of a scaling matrix.

Euclidean distance uses the identity matrix as the scaling matrix, implying that covariates are not scaled and are treated as equally weighted and uncorrelated. In contrast, scaled Euclidean distance defines S as the diagonal of the pooled covariance matrix, incorporating only the variances of the covariates to account for differences in scale. For Mahalanobis distance, S is the full pooled covariance matrix, allowing the metric to account for both the scale and correlations among covariates.² The robust Mahalanobis distance, on the other hand, is computed using the ranks of covariate values rather than their raw values, with an adjustment for tied ranks to enhance robustness to outliers and non-normality.³ Euclidean distance is often used as the default for continuous, standardized covariates, while the Mahalanobis distance is preferred when the covariates exhibit different scales or are correlated.

With these considerations, we decided to employ NNM using the scaled Euclidean distance.

¹ King, G., & Nielsen, R. (2019). Why Propensity Scores Should Not Be Used for Matching. Political Analysis, 27(4), 435–454. Retrieved December 5, 2025, from https://www.cambridge.org/core/journals/political-analysis/article/abs/why-propensity-scores-should-not-be-used-for-matching/94DDE7ED8E2A796B693096EB714BE68B.

² Rubin, D. B. (1980). Bias Reduction Using Mahalanobis-Metric Matching. Biometrics, 36(2), 293–298. Retrieved December 5, 2025, from https://www.jstor.org/stable/2529981.

³ Rosenbaum, P. R. (2010). Design of Observational Studies. Springer. Retrieved December 5, 2025, from https://www.stewartschultz.com/statistics/books/Design%20of%20observational%20studies.pdf.