Machine Learning for Alpha
Most machine learning in finance fails for one reason: the data is mostly noise and the researcher fools themselves. This course is about not being fooled — building features and labels that don't leak, validating in a way time series can't game, and proving an edge survives the hundreds of configurations you secretly tried.
ML done right in markets — financial features and labeling, the leakage trap and why ordinary k-fold fails on time series, purged and embargoed cross-validation, the multiple-testing backtest-overfitting problem and the deflated Sharpe ratio, tree ensembles versus linear models, MDI/MDA feature importance, and folding ML signals into the stat-arb pipeline. Overfitting is the enemy throughout.
Throw a random forest at a price series and it will hand you a backtest with a Sharpe of 4 and a smile. It is lying. Markets are the most adversarial dataset machine learning ever meets: the signal-to-noise ratio is brutally low, the data generating process drifts under your feet, the samples overlap and bleed information into each other, and — worst of all — you are in the loop, quietly trying configuration after configuration until something “works.” Almost every ML-in-finance failure traces back to the same root cause: the researcher fooled themselves, and the backtest was complicit.
This course is the antidote. It is not a tour of algorithms — you already know what a decision tree is, or you can look it up. It is a tour of the traps, and the discipline that defeats them. The single enemy, named on every page, is overfitting: mistaking noise for signal, in-sample for out-of-sample, luck for skill. Here is the arc, from raw data to a deployable signal:
- ML and the overfitting enemy — why markets break standard ML intuition: low signal-to-noise, non-stationarity, and the human-in-the-loop multiple-testing problem. The mindset shift from “find the best model” to “don’t get fooled.”
- Financial features & labeling — turning prices into stationary, informative features, and the labeling problem nobody warns you about: fixed-horizon labels and their flaws, the triple-barrier method, meta-labeling, and sample weighting for overlapping outcomes.
- Leakage & purged cross-validation — the leakage that makes a backtest glow, why ordinary k-fold is invalid on time series with overlapping labels, and the fix: purging, the embargo, walk-forward, and combinatorial purged CV.
- Backtest overfitting & the deflated Sharpe — the multiple-testing trap, why the best of N tried strategies is inflated by luck, the probability of backtest overfitting, the deflated Sharpe ratio, and the minimum backtest length you’d need to trust a result.
- Models, ensembles & feature importance — linear versus tree models on low-signal data, why regularization and ensembling (bagging, boosting) earn their keep, and how to read MDI and MDA feature importance without being misled by them.
- From ML signal to portfolio — folding a noisy ML forecast into the stat-arb pipeline: signal-to-position sizing, combining ML alphas with the breadth engine, respecting capacity and turnover, and the deployment discipline that keeps a live signal honest.
By the end you will treat a glittering backtest the way a forensic accountant treats a too-good quarter: with suspicion, a checklist, and the tools to find the fraud. A graded final exam runs the whole discipline back at you at the end, one locked question at a time.
In this topic
- 1 ML and the Overfitting Enemy Why markets break ordinary machine-learning intuition — punishingly low signal-to-noise, a non-stationary data-generating process, and a human-in-the-loop multiple-testing problem — and the mindset shift from 'find the best model' to 'don't get fooled.' 14 min
- 2 Financial Features & Labeling Turning prices into stationary, informative features without throwing away memory, and the labeling problem nobody warns you about: fixed-horizon labels and their flaws, the triple-barrier method, meta-labeling, and sample weighting for overlapping outcomes. 15 min
- 3 Leakage & Purged Cross-Validation The data leakage that makes a backtest glow, why ordinary k-fold cross-validation is invalid on time series with overlapping labels, and the fix that financial ML demands: purging, the embargo, walk-forward testing, and combinatorial purged cross-validation. 15 min
- 4 Backtest Overfitting & the Deflated Sharpe The multiple-testing trap that makes most published backtests false: why the best of many tried strategies is inflated by luck, how the expected maximum Sharpe under the null grows with the number of trials, and the tools that fight back — the probabilistic and deflated Sharpe ratios, the probability of backtest overfitting, and minimum backtest length. 16 min
- 5 Models, Ensembles & Feature Importance Choosing and taming models for low-signal data: linear models versus tree ensembles, why regularization and ensembling (bagging and boosting) earn their keep when the data is mostly noise, and how to read MDI and MDA feature importance without being misled by them. 15 min
- 6 From ML Signal to Portfolio Folding a noisy ML forecast into the statistical-arbitrage pipeline: turning a prediction into a position, combining ML alphas through the breadth engine, paying for fast signal decay with turnover and capacity, retraining against concept drift, and the deployment discipline that keeps a live signal honest. 15 min
- 7 Machine Learning for Alpha — Final Exam The graded final exam for Machine Learning for Alpha: low signal-to-noise and non-stationarity, financial features and triple-barrier labeling, data leakage and purged/embargoed cross-validation, backtest overfitting and the deflated Sharpe ratio, models, ensembles and MDI/MDA feature importance, and folding a noisy ML signal into the stat-arb pipeline. 18 min
Mark course as finished
Done with every lesson? Lock it in — your progress is saved on this device.