Feature Engineering
Feature engineering is the process of turning raw financial data into the input variables (called "features") that a machine learning model uses to make predictions. The quality of these features often matters more than the choice of model, because a model can only learn from the information the features actually contain.
In machine learning for finance, raw data rarely arrives in a form a model can use directly. Prices, volumes, filings, and economic releases must be transformed into clean, comparable, point-in-time variables. Feature engineering is the craft of designing those transformations: deciding what to measure, over what window, and how to scale it.
Definition
A "feature" is a single measurable input to a model, such as a stock's 12-month price change, its price-to-book ratio, or the change in trading volume relative to its recent average. Feature engineering is the work of constructing these inputs from raw observations so that the patterns relevant to the prediction task become easier for the model to detect.
The motivation is straightforward: most models cannot infer relationships that are not encoded somewhere in their inputs. A linear model given only raw prices cannot discover momentum on its own, but a feature equal to the trailing 12-month return hands that relationship to the model directly. Good features reduce the burden on the model and make the learned relationships easier to interpret.
Key Principle
A model can only learn from what its features encode. Garbage features produce garbage predictions regardless of how sophisticated the algorithm is. This is why practitioners often say that feature engineering, not model selection, drives most of the real difference in results.
Common Transformations
Financial features are usually built from a small set of recurring transformations. Each addresses a specific property of market data, such as differing scales, changing volatility, or skewed distributions.
| Transformation | What It Does | Why It Helps |
|---|---|---|
| Returns instead of prices | Converts raw price levels into percentage changes | Prices trend and are not comparable across stocks; returns are more stationary and comparable |
| Ratios and normalization | Divides one quantity by another (for example, valuation ratios) | Puts companies of different sizes on a common scale |
| Rolling windows | Computes averages, volatility, or sums over a trailing period | Captures recent behavior while smoothing single-day noise |
| Standardization (z-scores) | Centers a feature and scales it by its spread | Lets features with different units be compared and combined |
| Cross-sectional ranking | Ranks each stock against its peers at a point in time | Reduces the influence of outliers and shifting overall market levels |
Many of these transformations connect directly to documented effects. A trailing return feature encodes momentum, a valuation ratio encodes value, and a rolling standard deviation encodes recent risk. In this sense, feature engineering is where the insights of factor investing and the mechanics of machine learning meet.
Look-Ahead Bias and Data Leakage
The single most damaging error in financial feature engineering is using information that would not have been available at the time of the prediction. This is called look-ahead bias, and it produces features that quietly leak future knowledge into the model. The result is a backtest that looks impressive but cannot be repeated in live trading.
Leakage takes many forms. A common one is using a company's restated earnings rather than the figure that was actually reported on the original date. Another is computing a normalization statistic (such as a mean or standard deviation) over the full dataset, including future observations, before splitting the data for testing. Any statistic used to transform a feature must be computed only from information available up to that point. Guarding against leakage is closely tied to disciplined cross-validation and out-of-sample testing.
Known Limitations
Limitations to Keep in Mind
- More features can hurt, not help. Adding many candidate features increases the chance that a model fits noise rather than signal. Each additional feature gives the model another way to find a spurious pattern, which raises the risk of overfitting.
- Look-ahead bias is easy to introduce. Features that accidentally include future information inflate apparent accuracy. The error is often subtle and may only surface when live results fall far short of the backtest.
- Features decay. A transformation that captured a profitable pattern in the past may stop working as markets adapt and the underlying relationship weakens. Features require ongoing review, not a one-time design.
- Design choices embed assumptions. The choice of window length, scaling method, or ranking scheme reflects a judgment about how markets behave. Different reasonable choices can produce materially different models, so the apparent objectivity of a quantitative pipeline can be misleading.
- Engineering effort is hard to validate. Because analysts can try many feature variants, the temptation to keep the version that tested best is strong. This is a form of data snooping that requires careful cross-validation to control.
Applications
Feature engineering underpins most quantitative strategies that use statistical or machine learning models. In equity selection, features often encode established characteristics such as value, momentum, quality, and volatility, which makes the pipeline a natural bridge to multi-factor models. In risk modeling, engineered features summarize recent volatility regimes and correlations. In text-driven strategies, features are extracted from filings and news through natural language processing (NLP).
Across these applications, the same discipline applies. Features must be defined point-in-time, scaled sensibly, and validated honestly. The most advanced model cannot compensate for features that encode noise, leak the future, or no longer reflect how markets behave.
Further Reading
- Guyon, I. and Elisseeff, A. (2003). "An Introduction to Variable and Feature Selection." Journal of Machine Learning Research, 3, 1157–1182.
- Gu, S., Kelly, B. and Xiu, D. (2020). "Empirical Asset Pricing via Machine Learning." The Review of Financial Studies, 33(5), 2223–2273.
- Kuhn, M. and Johnson, K. (2019). Feature Engineering and Selection: A Practical Approach for Predictive Models. Chapman and Hall/CRC.
Related Terms
Foxholm Financial is a fee-only registered investment adviser serving Georgia. We bring quantitative rigor to every client engagement. Explore our services or get in touch to discuss how we can help. To see how this kind of analysis informs real client work, explore a Strategic Portfolio Review.
Are you an institution or FinTech firm? Learn about our Quantitative Consulting Services.
Foxholm Financial trains the next generation of quantitative analysts. Students and early-career researchers can explore our quantitative investment fellowships.
This content is for educational and informational purposes only and does not constitute an offer to sell or a solicitation of an offer to buy any securities. Nothing herein constitutes investment advice or recommendations tailored to your individual situation. All investments involve risk, including the potential loss of principal. Past performance is no guarantee of future results. Information presented is believed to be factual and up-to-date, but Foxholm Financial does not guarantee its accuracy and it should not be regarded as a complete analysis of the subjects discussed. Before making investment decisions, consult with a qualified financial advisor who can evaluate your specific circumstances.