This page documents the actual pipeline behind the model referenced throughout the pitch: how loan data is collected and labeled, how the classifier is built and validated, and how its output becomes an investment decision. It's written for technical diligence, not for the main narrative — hence its own page.
The training data comes from the fund's own historical activity on the IFC platform — loans already funded, each with a known result. A human authenticates interactively for every extraction run; the pipeline itself never stores or transmits the underlying credential, only the resulting session.
Extraction happens in two passes. The first pulls the portfolio list — deal terms and a profile summary for every historical loan. The second visits each loan's own detail view to backfill the fields the list doesn't expose: bureau score, the full delinquency-bucket breakdown, revolving and fixed-account balances and limits, and inquiry history. Each historical loan is then tagged by its real outcome — paid on time or defaulted — once that outcome is known, and the model retrains as newer loans mature into labeled data.
The 62 raw fields go through a fixed preprocessing pipeline before they reach the model: date fields (application date, employment start, bureau enrollment, most recent inquiry, oldest and newest account openings) convert to a numeric day-count; categorical fields (the platform's risk grade, loan purpose, housing status, car ownership, education level, employment type, address-change flag) one-hot encode; every numeric input is then standardized to zero mean and unit variance. That expands the 62 raw fields into 81 model inputs.
The classifier itself is a logistic regression, not an ensemble or a neural network. That's deliberate: with a labeled sample still in the low hundreds and 81 encoded inputs, a high-variance model would fit noise. Strong L2 regularization (C = 0.05) and a balanced class weight — correcting for the natural skew between paid and defaulted loans — keep it stable. The tradeoff pays off in two ways a regulated fund needs: every coefficient is auditable (defensible to LPs and to CNBV-adjacent diligence, unlike a black-box score), and inference is fast enough — well under two seconds per application — to run against an entire open listing with no serving infrastructure.
On each run, the pipeline pulls every open application the platform has already approved and published — loans that already passed the platform's own underwriting — and re-extracts the same full profile used in training. Each one runs through the identical preprocessing, reindexed to the exact 81 training columns, then through the trained model.
The output is a ranked list, highest to lowest repayment probability. The investment rule is a single, fixed threshold: only applications scoring above 85% qualify for capital. That's a second, independent screen layered on top of the platform's own risk grade — which is itself one of the 81 inputs, not a stand-alone gate. As more historical loans mature and their outcome becomes known, the labeled sample grows and the model is retrained and revalidated, so the threshold's calibration is checked against fresh data rather than fixed permanently at launch.
Current cross-validated performance: ROC-AUC 0.82 (5-fold). Trained on a manually labeled, growing sample of loans with known outcomes on the IFC platform. Revalidated as the sample grows. This page describes methodology, not implementation — specific scraping endpoints and credentials handling are omitted by design.