Feature Engineering for Financial ML: What We Use and Why
The 23 features that actually matter in our equity signal models. Includes rolling volatility ratios, order flow imbalance, and why we stopped using RSI.
After two years of iterating on our equity signal models, we have settled on 23 features that consistently contribute to prediction accuracy. This is not a theoretical list — these are the features running in production on trade today.
The most important category is volatility ratios: the ratio of short-term (5-day) to long-term (30-day) realised volatility. When this ratio spikes, it indicates a regime change. We compute this for both the target instrument and its sector ETF, giving us 4 features from this family alone.
Order flow imbalance is our second-strongest signal. We measure the ratio of market buys to market sells over rolling windows of 15 minutes, 1 hour, and 1 day. Persistent imbalance in one direction precedes price moves more reliably than price momentum alone.
What we dropped: RSI, MACD, and most traditional technical indicators. In our testing, these added no predictive power on top of our engineered features. They are lagging indicators derived from the same price data our models already see. Including them just added noise.
Cross-asset features round out our set. We track correlation rolling windows between the target instrument and 5 related assets (sector ETF, market index, volatility index, currency pair, and commodity). Correlation breakdowns are strong predictors of upcoming moves, especially in stress periods.