Machine Learning in Finance

This two-day course gives finance professionals a practical, end-to-end introduction to machine learning for market prediction. Participants learn how to prepare data, build leak-free features, train time-series-aware models, evaluate them properly, and convert model outputs into simple trading signals and backtests.

• Portfolio managers
• Traders
• Analysts and quants
• Risk and model-validation staff
• Anyone looking to apply ML in a practical investment context

• Frame ML problems correctly for financial data
• Build clean features and targets without leakage
• Train and compare baseline and tree-based models using temporal CV
• Use AutoML as a benchmark
• Explain model drivers with SHAP
• Turn predictions into positions and run simple backtests

• Basic Python
• Familiarity with financial markets and return data
• All code and templates provided; no ML background required

Foundations & Strong Baselines

Data → Features → Targets

Theory

• Problem framing: regression vs classification vs ranking; target definitions that survive costs
• Data hygiene for finance: calendars, mixed frequencies, forward-fill rules, leakage traps
• Feature playbook: returns/log-returns, lags, rolling stats/z-scores, momentum/carry/spreads
• Time-aware splits: temporal CV vs random; baseline benchmarks

Practical

• Build a leak-free feature set; create 5-day return (regression) + up/down (classification) targets
• Export a clean training dataframe + quick EDA plots

Supervised ML for Tabular Time Series

Theory

• Models: Linear/Logistic (with regularization), Random Forest, Gradient Boosting, XGBoost/LightGBM
• Metrics: RMSE/MAE, PR-AUC vs ROC-AUC (imbalance), calibration, regime stability checks
• Sensible tuning: early stopping, simple grids; when to keep it simple

Practical

• Train ridge/elastic net vs LightGBM using temporal CV helpers
• Compare performance + calibration; record a 1-page “model card”

Robustness, AutoML & From Signals to Trades

Walk-Forward, Robustness & AutoML Lab

Theory

• Walk-forward validation, gap splits, regime slicing; drift checks
• Quantile regression (pinball loss) vs probability forecasts; when each helps

Practical

• Run a lightweight AutoML pass (e.g., scikit-learn-compatible AutoML) on the Session-1 dataset with time-aware CV
• Compare AutoML’s top model vs your manual LightGBM from Day 1 (metrics, stability, training time)
• Optional: quick quantile regression run to estimate upside/downside bands

Explainability → Policy Rules → Simple Backtest

Theory

• Explainability for PMs/validators: SHAP (global/local), permutation importance, sanity checks
• From predictions to trades: thresholds, confidence-based sizing, smoothing, turnover controls
• Backtest basics: costs, slippage, drawdown, turnover; common pitfalls/leakage

Practical

• SHAP on best model; verify key drivers + sign consistency
• Convert scores to positions (calibrated threshold + size by confidence / quantile bands)
• Run vectorized daily backtest; report equity curve, IR, drawdown, turnover sensitivity to costs