Machine learning (ML) is a family of methods that use data or experience to fit a model, discover structure or improve a decision rule under a chosen learning objective. In quantitative finance, that can mean estimating a future outcome, grouping assets by their behaviour or learning an execution policy. Successful fitting is the beginning of the evidence: it does not by itself establish a useful forecast, a causal explanation or a profitable strategy.
The researcher still chooses what the system sees, what it is asked to learn and how success will be judged. A complex algorithm cannot repair a target that answers the wrong economic question.
What supplies the learning signal?
| Approach | What guides learning | Possible financial task |
|---|---|---|
| Supervised learning | Examples paired with observed targets | Predict a numerical outcome with regression, or a category with classification |
| Unsupervised learning | Structure in observations, without a supplied target for each example | Group instruments with similar return behaviour |
| Reinforcement learning | Rewards from actions in an environment | Learn a trading or execution policy under specified costs and constraints |
These are learning setups, rather than a ranking of model quality. Labels are central to supervised learning; they are not a requirement for every form of ML. A reinforcement learner's result depends on its environment and reward design: optimising a simulator does not establish performance in a live market.
Statistical methods and ML overlap substantially. A fitted linear regression belongs in this landscape just as a neural network does. “Learning” does not require human-like reasoning, and a larger model is not automatically a stronger baseline.
A model that learns only one number
Take two synthetic, dimensionless training targets: 1 and 3. Restrict the model to predicting the same constant for every observation. Choose mean squared error as the fitting objective:
Because the squared term cannot be negative, the unique minimum is at . The learned prediction is the sample mean. Its errors against the two training targets are and , giving a mean squared training error of 1.
Original synthetic dimensionless constant-fit example; the minimum is a training result, not evidence of financial predictive skill.
This tiny model shows the mechanism: data, an allowed family of models and an objective determine a fitted result. It has no input-dependent predictive rule, but it can serve as a simple benchmark for a richer regression. The training error says how well it fitted those examples. It says nothing yet about its error on later observations.
The financial test comes after the fit
For a supervised financial model, distinguish the features available at decision time from the target that may only become known later. A future return can be a training target once its outcome is available; putting that future information into a historical input creates look-ahead bias. Preprocessing can leak information too, such as scaling the past using a full-history estimate.
Use development data to fit and select the procedure, then evaluate it on separate later data. Freeze the rules for feature construction, tuning and scheduled refitting before opening that final evaluation. Repeatedly changing the model after viewing the same test makes those observations part of development. Overfitting can occur even when every timestamp is correct.
The owned lesson Time-Ordered Train, Validation and Test Design develops this distinction through training eligibility and walk-forward schedules. A public preview is available; the full lesson requires membership.
A model can predict a statistical association without identifying what would happen under an intervention. It can also predict an outcome more accurately while producing worse trading results. Moving from forecast to strategy adds position sizing, execution, transaction costs, financing and risk constraints. Compare both predictive performance and the resulting economic decisions against a credible simpler baseline, using information and costs available at the intended trading times.
Sources
- Hastie, T., Tibshirani, R. and Friedman, J., The Elements of Statistical Learning, 2nd ed., corrected 12th printing (January 2017), §§2.1–2.3.1, pp. 9–12, and §14.1, pp. 485–487. Local grounding copy read for supervised learning, least squares and unsupervised learning.
- Google for Developers, What is Machine Learning?, learning setups, including reward-based learning. Checked 27 October 2026.
- López de Prado, M. (2018), Advances in Financial Machine Learning, §§11.2–11.4. Local grounding copy read for historical simulation limits, selection, costs and the distinction between predictive features and monetisation. The constant-fit example and figure here are original.