Structural vulnerability versus timing of deterioration: a prospective evaluation of supervisory early warning for MSE loan portfolios in Peruvian banks and municipal savings banks, 2018–2026

OverviewThis repository hosts the replication dataset, pre-specified empirical protocols, frozen predictive models, and reproducible analytical pipeline for the research manuscript: "Structural vulnerability versus timing of deterioration: a prospective evaluation of supervisory early warning for MSE loan portfolios in Peruvian banks and municipal savings banks, 2018–2026" by Renzo Alberto Taco Coayla, Martha Judith Paredes Vignola, and Elard Ricardo Rodriguez Marca. The study evaluates whether early warning systems (EWS) applied to panel data distinguish structural institutional vulnerability ("who is at risk") from the timing of deterioration ("when deterioration occurs") across micro and small enterprise (MSE) credit portfolios. Dataset Characteristics and Sample Construction- Time Span: Monthly panel from January 2018 to September 2026 (1,801 institution-month observations; 1,155 downloaded primary files).- Target Institutions: 25 regulated financial institutions in Peru operating an active MSE loan portfolio of at least PEN 50 million in the alert month (13 commercial banks, 11 Cajas Municipales de Ahorro y Crédito - CMACs, and 1 Caja Municipal de Crédito Popular - CMCP).- Primary Data Source: Official statistical supervisory reports published by the Superintendencia de Banca, Seguros y AFP (SBS, Peru): Direct loans by type and status (B-2334, C-1228), delinquency by modality (B-2362, C-1242), and financial ratios (B-2401, C-1301).- Prediction Target (Event): A 3-month increase in the MSE non-performing loan (NPL) ratio exceeding the 85th percentile of the embargoed training distribution (ΔNPL t+3 ≥ 1.03 percentage points). Includes secondary targets for relative deterioration and high-risk portfolios (past-due plus refinanced loans).- Feature Space: 27 predictor variables built with calendar lags (t−1, t−3, t−6, t−12) covering portfolio scale, growth, NPL dynamics, refinancing shares, ROA, liquidity, leverage, provisions, and institutional profiles. Repository Contents- data/: Clean panel data files in tidy tabular formats (.csv and .rds) covering all 1,801 observations across estimation, validation, test, and prospective confirmation splits.- models/: Serialized frozen predictive models (Random Forest Layer A, Full Logit, and NPL-dynamics Logit) along with their respective SHA-256 integrity hashes.- protocols/: Pre-registered, version-controlled execution protocols specifying sample boundaries, embargo structures, and formal hypotheses (H1–H5).- scripts/: End-to-end reproducible pipeline routines (scripts numbered 10 to 42) for data extraction, variable computation, between/within AUROC decomposition, cluster-robust bootstrapping (1,000 institution-level replicates), TreeSHAP explainability, and exposure-weighted evaluation. Software RequirementsAll routines were developed and tested using standard statistical computing environments (Python / R). Requirements and library dependencies are detailed within the repository readme files. License and UsageThis dataset and the accompanying code are licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. Derivative works and academic citations should refer to this repository via its registered DOI and cite the corresponding research article.