Machine Learning for Trading

At a glance

Signal-to-noise: Extremely low compared with typical ML domains
Stationarity: Absent; the data-generating process changes
Best applications: Signal combination, execution, risk estimation
Worst application: Predicting price direction from raw price history

Key takeaways

Machine learning excels where signal is strong and the underlying process is stable. Financial markets are the opposite on both counts.
The most valuable applications are combining many weak signals, estimating risk, and improving execution, not forecasting direction.
Standard cross-validation is invalid on time series: random splits leak information across adjacent, correlated observations.
Labelling is a modelling decision. Fixed-horizon returns ignore path and volatility; methods that account for both produce more meaningful targets.
Simple models with few features and honest validation consistently outperform complex models that have been validated carelessly.

Why financial data is a difficult case

Property	Typical ML domain	Financial markets
Signal-to-noise ratio	High: images, speech	Extremely low; mostly noise
Stationarity	Stable relationships	Regimes change; relationships decay
Sample size	Millions of independent examples	Thousands of overlapping, correlated observations
Feedback effects	None: a cat does not react to the classifier	Strong: profitable predictions change the market
Label clarity	Unambiguous	A modelling choice with many defensible options
Cost of error	Symmetric	Asymmetric; tail losses dominate
Validation	Random cross-validation works	Random splits leak; requires purged time-series methods

Where machine learning genuinely helps

Combining many weak signals. Given twenty features each with a small information coefficient, learning how to weight them is a well-posed problem that linear and tree-based models handle well.
Execution optimisation. Predicting short-term price impact and optimal order placement uses high-frequency data where sample sizes are large and the target is more directly observable.
Risk and volatility estimation. Volatility is far more predictable than returns, which makes it a much better target than direction.
Regime classification. Identifying which market state you are in, and therefore which strategy to enable, is a lower-dimensional and more tractable problem than forecasting.
Alternative data extraction. Converting text, images, or unstructured data into a usable numeric signal is a genuine ML task where the models excel.
Anomaly detection. Flagging data errors, unusual market conditions, or system malfunctions, which supports the trading process rather than making the trades.

The pattern is consistent: machine learning contributes most where the target is something other than the direction of price, and where the sample is large relative to the number of parameters.

Validation that actually works

1
Never use random cross-validation
Adjacent observations in a time series are correlated, so random splits place nearly identical examples in both training and test sets, producing wildly optimistic scores.
2
Use time-series splits with purging
Train on earlier data and test on later data, and remove observations near the boundary whose label horizons overlap the training period.
3
Add an embargo period
Skip a gap after the training set before the test set begins, to prevent serial correlation from leaking across the boundary.
4
Account for overlapping labels
If each label spans five days, consecutive observations share information and are not independent. Sample weighting or non-overlapping sampling is required.
5
Validate the trading result, not the classification score
A model with 53 percent accuracy can be highly profitable, and one with 58 percent accuracy can lose money if it is wrong on the largest moves.
6
Deflate for the number of models tried
Hyperparameter search is multiple testing at scale. A hundred configurations means the best one is an extreme draw regardless of merit.

Labelling: the underrated decision

What you ask the model to predict determines what it learns. Fixed-horizon returns are the default choice and frequently the wrong one.

Labelling method	Description	Problem it solves
Fixed-horizon return	Return over the next N bars	Simple; ignores path and volatility entirely
Volatility-scaled return	Return divided by recent volatility	Makes labels comparable across regimes
Triple barrier	Which of profit target, stop, or time limit is hit first	Reflects how the position would actually be managed
Meta-labelling	Predict whether an existing signal will be correct	Separates direction from position sizing
Trend scanning	Label by the statistical significance of the local trend	Reduces noise in the target

The triple barrier method deserves particular attention because it aligns the label with the trade you would actually take, including the stop. A model trained on fixed-horizon returns can be confidently right about the endpoint while every trade is stopped out along the way.

Practical guidance

Start with a linear model. If a regularised linear model finds nothing, a neural network will find noise. Linear models also make it obvious what the model is using.
Use few features with economic rationale. Each additional feature increases the search space and the chance of fitting. Twenty well-motivated features beats two hundred generated ones.
Predict something other than direction where possible. Volatility, regime, or whether an existing signal will work are all easier targets.
Use meta-labelling. Let a simple rule generate signals and use the model to decide which to take and how large. This is one of the more reliable applications.
Validate with purged, embargoed time-series splits. Anything else on financial data produces results that will not survive.
Simulate trading, with costs, as the final evaluation. Never promote a model on a classification metric.
Expect decay. Monitor live performance against the validated distribution and retire models that fall outside it, exactly as with any other strategy.

Frequently asked questions

Can machine learning predict stock prices?

Not reliably in the sense usually intended. Price direction has an extremely low signal-to-noise ratio and a non-stationary relationship to any feature set. Machine learning contributes more where the target is volatility, regime, execution quality, or the reliability of an existing signal, all of which are more predictable than direction.

Why does my ML model work in backtest and fail live?

Most commonly because validation used random cross-validation, which leaks information across correlated adjacent observations, or because hyperparameter search over many configurations produced an extreme draw. Overlapping labels and full-sample feature normalisation are the other frequent causes.

What is meta-labelling?

Using a simple rule to generate candidate trades, then training a model to predict which of those trades will be profitable. It separates the direction decision from the selection and sizing decision, which is a more tractable learning problem and tends to produce more robust results than predicting direction directly.

Should I use deep learning for trading?

Rarely, at retail scale. Deep learning requires large samples relative to its parameter count, and financial data provides few genuinely independent observations. It has legitimate applications in extracting signal from unstructured alternative data, but for tabular price and fundamental features, regularised linear and tree-based models usually perform better and are far easier to validate.

How many features should a trading model use?

Fewer than instinct suggests. Ten to thirty features with clear economic rationale is a reasonable range for most retail applications. Each additional feature expands the search space and increases the probability of fitting noise, and financial datasets rarely contain enough independent observations to support large feature sets.

Test this idea before you trade it

Describe the rules in plain language and AlgoTrader AI turns them into a structured strategy blueprint with a configurable historical backtest, cost assumptions, and exportable code.

Build a backtest

Keep reading

Referenced by

Educational use only. This guide explains how a strategy works. It is not investment advice, not a recommendation, and no result described here is a forecast. Test any approach on historical and out-of-sample data, size positions conservatively, and never risk money you cannot afford to lose.