Alternative behavioral data as screening signals: incremental predictive power and economic implications in credit scoring among thin-file borrowers in Vietnam

Student: Võ Ngọc Tường Vy, VNP-29

Supervisor: Dr. Trương Đăng Thụy

Abstract:

This thesis examines whether alternative behavioural-information proxies provide incremental predictive power beyond traditional credit-information proxies in screening small-value, short-term consumer loans for thin-file and new-to-bank applicants in Vietnam. The anonymised and feature-masked data-set was provided by KCI Credit Information Joint Stock Company and contains 10,023 unique applicant-originated-loan records. Predictors were measured at or before the scoring point, while the dependent variable records whether the corresponding one-month loan subsequently reached 30 days past due (DPD30+) after being observed through maturity and for at least 30 additional days. The primary analysis uses unweighted Baseline and Extended Logistic Regression models. RF52 uses the same final 52 predictors as the Extended Logistic Regression and provides a clean algorithm-only robustness comparison. A separate supplementary RF92
model uses all 92 screened source predictors to assess the broader KCI information environment; it is not used to identify the incremental contribution of the behavioural block.

After diagnostic removal of one duplicated traditional predictor and twelve deterministically dependent behavioural predictors, the Baseline model retained 18 predictors and the Extended model retained 52 predictors. On the ordered out-of-time test sample, the Extended Logistic Regression increased AUC from 0.6515 to 0.6954 and KS from 0.2218 to 0.2999; the paired DeLong test confirmed that the AUC gain was statistically significant (z = 4.934, p < 0.001). Within the thin-file proxy segment, AUC increased from 0.5699 to 0.6449 and KS from 0.1300 to 0.2516. The Extended model also produced a lower Brier score and lower observed bad rates among approved loans at the evaluated predicted-risk cut-offs. Weighted Logistic Regression produced similar ranking results but was retained only as a supplementary sensitivity analysis. The findings indicate that governed behavioural information can complement traditional credit information,
particularly where formal credit histories are sparse; they do not imply that alternative data should replace conventional credit assessment or that the estimated probabilities constitute
regulatory PD measures.

Keywords: alternative behavioural information; credit scoring; DPD30+; thin-file borrowers; new-to-bank; logistic regression; Vietnam; model governance