Feature Engineering

In trading machine learning, the choice of inputs usually matters more than the choice of model. Feature engineering means turning raw data, such as prices, volumes, financial statements or news, into variables that capture something meaningful and that a model can learn from. Good features are grounded in a reason the market might behave a certain way, are calculated using only past information and are scaled so they mean the same thing across time and across assets. A simple model with thoughtful features almost always beats a complex model fed raw prices.

Main feature families#

Make features stationary#

Raw prices trend and change scale over time: a stock at $20 in 2015 and $200 today. A model that learns from price levels learns nothing transferable. Convert to forms with stable statistical properties:

• Returns instead of prices.
• Ratios such as price divided by its 50 day average.
• Changes in rates or spreads.
• Volatility scaled values, such as a return divided by recent volatility.

See Stationarity, Differencing and Unit Roots.

Make features comparable#

Use only past data for any scaling. A z score computed with a full sample mean leaks the future. See Rolling and Expanding Windows.

Avoiding leakage#

See Data Leakage and Look-Ahead Bias.

How many features?#

More features are not always better. Each extra feature adds a chance of fitting noise, and many features are highly correlated. Start with a few well motivated ones, check their individual predictive power out of sample, and add more only if they improve validation results. Techniques such as principal component analysis or clustering can group correlated features. See Overfitting and Curve Fitting.

Testing a single feature#

Before modelling, test each feature alone: sort assets or periods into groups by the feature and compare later returns, or compute the information coefficient, the rank correlation between the feature and future returns. A feature that shows no relationship alone rarely becomes useful inside a model. See Signal Discovery.

Frequently asked questions#

What is feature engineering in trading?#

Creating model inputs from raw data, such as returns, volatility, valuation ratios or sentiment scores, in forms a model can learn from.

Why should features be stationary?#

Non stationary inputs such as raw prices change scale over time, so patterns learned in one period do not apply in another.

How do I avoid look ahead bias in features?#

Use only data available before each prediction, shift features appropriately, use rolling statistics and use point in time fundamental data.

Next, learn how to validate models properly in Model Evaluation and Cross-Validation.