Track 02 · 8 modules
Data Science
Practical machine learning, statistics, and modeling for credit and risk problems — from EDA gotchas and feature engineering through backtesting, monitoring, Bayesian methods, and causal inference.
- 01
EDA for Financial Data — The Gotchas
Six ways financial data trips up exploratory data analysis: fat tails, survivorship bias, look-ahead bias, regime changes, missingness that means something, and time-zone landmines.
- 02
Feature Engineering for Credit & Risk
Ratios, lags, rolling windows, trend features, and WOE encoding — how raw loan data becomes model food without leaking the future or blowing up in production.
- 03
Time Series for Finance: Stationarity, ARIMA, GARCH
The three ideas that keep financial time-series work honest: know whether your series mean-reverts, model levels and volatility as different problems, and never trust an R² on two trending series.
- 04
Classification Metrics Done Right
AUC, Gini, and KS are the same family; accuracy lies whenever defaults are rare; precision-recall is what imbalance actually feels like; and calibration is a separate virtue your ranking metric cannot see.
- 05
Backtesting & Walk-Forward Validation
Random cross-validation quietly trains on the future. This module covers the honest alternatives — walk-forward schemes, purging and embargoes for overlapping labels, and the multiple-testing arithmetic that makes most backtests too good to be true.
- 06
Production ML for Risk: Monitoring, Drift, and Retraining
A deployed risk model starts dying the day it ships. What to watch (PSI, Gini trend, calibration, overrides), why credit's outcome lag forces you to monitor inputs, and how to retrain without chasing noise.
- 07
Bayesian Methods for Risk
What do you do with a portfolio of 300 loans and zero defaults? Frequentist statistics shrugs; Bayesian methods answer. Priors as reference classes, beta-binomial machinery, and shrinkage across segments.
- 08
Causal Inference in Lending
Prediction answers 'who will default?' Policy asks 'what happens if we change the rate, the limit, the cutoff?' Confusing the two is how correlation-only models embarrass you — in committee and in front of regulators.