ML × Quant Trading
8 articles ~1.2 h
- Alpha Decay and Feature Half-Life Every alpha factor decays. The half-life of an alpha's predictive power tells you how often to retrain, how much turnover to accept, and when to give up.
- Cross-Validation in Time Series Random k-fold produces numbers with no relation to live performance. Walk-forward, purge, embargo, and CPCV in full, plus a worked example that gets both gap widths right.
- Embedding Methods for Assets An embedding represents an asset as a vector of numbers that captures its relationships to other assets. This is how you feed 'Apple' or 'Goldman Sachs' into a machine learning model.
- Feature Engineering for Quant The algorithm is interchangeable; the features are the strategy. Three families of raw material, four tests a feature has to pass, five mistakes everyone makes once, and the full rebuild of naive momentum into a cross-sectional z-score.
- ML for Return Prediction Backtest Sharpe, live Sharpe, and deployable Sharpe are three different numbers. Where the leakage happens between them, how to define the target, what capacity takes back, and an honest range for the result.
- ML Trading Strategy Lifecycle The six stages from research to retirement, with the honest ratio, evaluation window, and exit condition for each, and why most strategies die on process rather than on modeling.
- Neural Networks for Time Series Neural networks for time series promise to learn temporal patterns from raw returns. In practice, transformers are not a magic bullet and only a few architectures work reliably.
- Reinforcement Learning for Portfolios Reinforcement learning trains an agent to make sequential decisions by reward and penalty. Applied to portfolios, it sounds perfect. The reality is more complicated.