Machine Learning for Alpha — Final Exam

This is the capstone. Six lessons taught the brutal craft of pointing machine learning at financial markets — where the signal-to-noise ratio is so vile that off-the-shelf ML, applied naively, produces nothing but confident, well-cross-validated, completely fictitious alpha. You learned why finance breaks the usual ML assumptions: the signal-to-noise is tiny (IC ≈ 0.02–0.05, R² near zero), the data-generating process is non-stationary (regimes shift, edges decay, concepts drift), and the analyst’s own repeated experiments make the whole exercise a multiple-testing minefield. You learned to build stationary-yet-memory-preserving features with fractional differentiation and to label outcomes the right way with the triple-barrier method and meta-labeling instead of naive fixed-horizon returns. You learned the seven faces of data leakage and why ordinary k-fold cross-validation is a lie on time series — and how purging and embargoing (and combinatorial purged CV) repair it. You learned why the best of N backtests is a near-guaranteed mirage, how the expected maximum Sharpe under the null grows like √(2 ln N), and how the deflated Sharpe ratio and probability of backtest overfitting put a real hurdle under a result. You learned the bias–variance logic behind regularization, why bagging kills variance while boosting chases bias (and overfits noise), and why MDI lies in-sample while MDA tells the out-of-sample truth — and that importance is never causation. Finally, you learned to fold a weak ML prediction into a real book: prediction → position, combining orthogonal alphas via IR ≈ IC·√breadth, respecting capacity, turnover and decay, and deploying with a locked test set, paper trading, live-IC monitoring and a kill-switch. No formula sheet, no hints, no take-backs: every answer locks the instant you submit, the wrong options are the exact traps that fool real desks, and your score stays hidden until the end.