<![CDATA[Default prediction models for peer-to-peer lending are often tested on random data splits and judged by accuracy, which can hide temporal leakage and says little about investor returns. This study evaluates a random forest on 1,020,768 resolved 36-month Lending Club loans issued from 2007 to 2018, using 30 walk-forward windows with a 36-month gap between validation and test periods and only information available when a loan is funded. A decision threshold chosen by validation return on investment (ROI) was compared with the default threshold of 0.5, lending to every loan a grade-based rule and logistic regression, with full and downsampled training data, five random seeds, and paired tests across windows. The random forest reached a mean AUC of 63.7–64.6%, close to logistic regression (64.2–65.0%). Overall windows, the ROI-based threshold raised mean window ROI from 4.04% (lend-all) to 4.97% (difference 0.92 percentage points, 95% CI −0.29 to 2.23) while lending 51% of the capital; no difference remained significant after Holm adjustment. The share of test loans resolved by the end of the data fell from over 99% in windows 1–19 to 7% in window 30, so the negative returns of late windows reflect censoring. In the 19 matured windows, lend-all earned 9.08% on average and the random forest did not improve on it: ROI-based thresholds lowered mean ROI by 0.92–1.15 percentage points and profit by 40–54%. Return-based evaluations of default models therefore need time-ordered tests that account for unresolved loans.]]>
Copyrights © 2026
