ML-Driven Strategies

Read these first

1. 01Building an Alternative-Data ML Pipeline

   Turning a raw feed of credit card transactions or satellite images into a usable trading signal takes a whole pipeline, cleaning, entity resolution, feature construction, and point-in-time-correct backtesting, and most of the effort in an alt-data strategy is in that pipeline, not in the final model.

2. 02Turning a Model Prediction Into a Position

   A model that outputs '0.3% expected return over 5 days' hasn't told you how many shares to buy, getting from a prediction to an actual position means scaling by conviction, volatility, correlation with existing holdings, and the cost of trading, and most of the money in a systematic strategy is made or lost in that translation step.

3. 03Why ML Strategies Degrade Live

   A backtest Sharpe ratio of 2.5 turning into a live Sharpe ratio of 0.5 is the single most common story in systematic trading, and it has a short list of usual suspects: overfitting to noise, subtle look-ahead bias, a training period that no longer resembles the live market, and crowding from everyone else who found the same signal.

4. 04Deep Models on Limit Order Book Data

   Feeding a neural network the raw sequence of bid and ask prices and sizes, tick by tick, lets it learn short-horizon price-move patterns a hand-built feature never captures, but the resulting model has to make a prediction and act on it faster than the exchange can invalidate it.

5. 05Ensembling Alpha Models

   Combining several imperfect return-prediction models into one blended signal, weighting each by its own skill and by how little it overlaps with the others, reliably beats picking the single best model, because the errors of independent models cancel in a way their individual signals never could.

6. 06Tree-Based Cross-Sectional Return Models

   Gradient-boosted trees rank thousands of stocks by predicted forward return using dozens of interacting features at once, capturing 'value is good, but only in low-momentum stocks' style non-linearities a linear factor model can't see, at the cost of being much easier to overfit.

Then the rest

1. 07Choosing the Target: Raw, Excess or Residual Returns
2. 08Crowding in ML-Driven Signals
3. 09Designing a Cross-Sectional Feature Set
4. 10Kill Switches and Human Overrides for ML Models
5. 11LLM-Generated Features for Trading Signals
6. 12Microstructure Features for Short-Horizon Models
7. 13Neutralising a Model's Output Against Risk Factors
8. 14Nonlinear Feature Interactions That Actually Pay
9. 15Online vs Batch Updates for Live Models
10. 16Panel Drift and Vendor Methodology Changes
11. 17Putting an NLP Signal Into Production
12. 18Putting Trading Costs Into the Training Objective
13. 19Regime-Conditional ML Models
14. 20Sequence Models for Multi-Horizon Return Forecasts
15. 21Short-Horizon ML Alpha Under Cost Constraints
16. 22Turnover Control in ML Strategies
17. 23Warning Signs of an Overfit ML Strategy
18. 24When an ML Signal Becomes a Strategy
19. 25Monitoring Input Feature Distributions
20. 26Ranking vs Regression Objectives for Alpha