Feature Engineering
Feature engineering is the process of transforming raw data into informative input variables that a machine-learning model can use to make better predictions.
Feature engineering is the practice of turning raw, messy data into the structured inputs a model actually learns from. In finance, raw market data is just a stream of prices and volumes; on its own that is rarely the most useful form. Feature engineering converts it into derived signals such as returns over various horizons, rolling volatility, momentum and mean-reversion measures, technical indicators (RSI, MACD, Bollinger Bands), volume profiles, calendar effects, and relationships between an asset and broader factors. The goal is to encode the patterns a human analyst would look for into numbers a model can weigh.
The quality of features usually matters more than the choice of algorithm. A sophisticated model fed weak inputs will underperform a simple model fed well-chosen ones. Good feature engineering also requires discipline about timing: every feature must be computable using only information available at the moment of prediction, or the model inherits look-ahead bias and backtests look far better than live results ever will. Features must also be reasonably stable and not so numerous that the model memorizes noise rather than learning signal, which is why feature selection and regularization often go hand in hand with feature construction.
For investors, feature engineering is where domain knowledge enters a quantitative system. It is the bridge between financial intuition (what tends to move a stock) and the statistical machinery that scores those ideas at scale. Poorly chosen features produce models that look clever but generalize badly; carefully chosen, leak-free features are what let a model find durable structure in noisy markets.
hedgewing.ai's models are built on roughly 45 engineered features spanning price action, volatility, momentum, volume, and cross-asset and factor relationships. These features feed a four-model deep-learning ensemble (LSTM, GRU, TCN, and Transformer), and each one is constructed to be strictly point-in-time so that nightly walk-forward backtests reflect only information that would have been available in real time.
Related terms
Supervised Learning · Look-Ahead Bias · Overfitting · Technical Analysis
Back to the hedgewing.ai glossary · See AI stock forecasts