ML for Alpha Research

Read these first

1. 01Blending Alpha Signals with Machine Learning

   Combining ten weak, mostly-independent signals into one score can beat any of them alone, but only if the blend is built to resist the temptation to let a flexible model quietly re-fit noise across the very signals it's supposed to be averaging out.

2. 02Building a Purged Walk-Forward Harness

   Ordinary cross-validation shuffles data into folds at random, which is exactly wrong for market data, because a model can peek at overlapping information between adjacent 'train' and 'test' rows unless the harness deliberately purges and embargoes around the split.

3. 03Designing a Reusable Signal Library

   A signal library is the difference between every researcher reinventing feature code from scratch and a desk that can test a new idea in an afternoon, but only if the library enforces the same point-in-time and versioning discipline every individual signal needs.

4. 04Framing Alpha Research as a Prediction Problem

   Turning a trading idea into something a machine learning model can be trained on requires deciding exactly what to predict, over what horizon, and how to score success, and the decisions that make a good academic prediction problem can quietly make a bad trading strategy.

5. 05Overfitting the Cross-Section

   A model tested across thousands of stocks at once can look robust simply because it has thousands of rows per day, but every one of those rows shares the same day's market-wide noise, so the model can be fitting a handful of lucky days, not a real cross-sectional pattern.

6. 06Regime-Conditional Alpha Models

   A single model trained across bull markets, crashes and grinding sideways periods learns an average behaviour that fits none of them well, regime-conditional models try to let the model behave differently in different market states, at the real cost of splitting an already-scarce sample even further.

Then the rest

1. 07Lagging Features to Enforce Point-in-Time Safety
2. 08Beating the Linear Baseline: When ML Actually Helps
3. 09Binary vs Continuous Labels for Alpha Models
4. 10Capacity of Machine-Learning Signals
5. 11Choosing the Prediction Target: Return, Rank or Direction
6. 12Clustered Feature Importance and Mean Decrease Accuracy
7. 13Conditioning Variables and Interaction Features
8. 14Converting Model Scores into Positions
9. 15Cost-Aware Trade Thresholds and No-Trade Bands
10. 16Deriving Regime Labels for Model Switching
11. 17Feature Half-Life and Smoothing Choices
12. 18Hyperparameter Search Budget and Multiple Testing
13. 19Label Overlap and Effective Sample Size
14. 20Leakage Audits for Alpha Pipelines
15. 21Matching Forward-Return Horizon to Signal Decay
16. 22Monitoring Live Alpha Model Decay
17. 23Orthogonalising Signals Before Blending
18. 24Pooled vs Per-Asset Models on Panel Data
19. 25Rank Targets and Cross-Sectional Normalisation
20. 26Refit Frequency for Cross-Sectional Models
21. 27Residualised Return Targets After a Risk Model
22. 28Train-Test Decay Curves for ML Signals
23. 29Turnover Penalties in Model Training
24. 30Volatility-Scaled Return Targets