Features or deep learning: what actually works on market data

There is a reliable pattern in what wins on tabular financial data, and it disappoints people who expect the answer to be a neural network.

The signal-to-noise problem

In image recognition, the label is essentially deterministic given the input — a cat is a cat. In market prediction, the same feature vector can precede a rise or a fall, and the best achievable accuracy is barely above chance.

This changes what model capacity buys you. With high signal-to-noise, more capacity extracts more structure. With low signal-to-noise, more capacity mostly extracts more noise, and it does so with great confidence.

Why boosted trees keep winning

• Sample efficiency. Useful results from tens of thousands of examples, not millions. Financial history is short — there are only so many market days.
• Robustness to scale and outliers. Trees split on order, not magnitude. A 10-sigma print does not distort the model the way it distorts a gradient.
• Interpretability. Feature importance and partial dependence are directly readable, which matters when an operator has to sign off on a live system.
• Fast iteration. Minutes to train means dozens of honest experiments per day. With low signal-to-noise, iteration count is the binding constraint.
• Missing values. Handled natively. Market data has gaps.

This is why our own library, TideGBM, is a gradient booster rather than anything more fashionable — and why the work went into purged validation rather than into capacity.

Where the effort actually pays

Not in architecture search. In features. A well-constructed feature encodes domain knowledge the model would otherwise need enormous data to discover — and the data does not exist.

Our pump-fade model runs 59 features in four families, and the honest breakdown of where the predictive power lives is instructive:

We publish that last row rather than hiding it. The formulas are the part that looks clever and contributes least, which is a common shape and worth expecting.

Where deep learning does win

Three narrow cases, all of which share a property: the input has structure a tree cannot exploit.

1. Raw order book sequences. Convolutional or attention models over the full book across time can find patterns that hand-built summary features flatten away — given a lot of data and event-resolution recording.
2. Cross-instrument structure. Learning relationships across hundreds of symbols simultaneously, where the interaction structure is the point.
3. Representation learning for downstream use. Learning an embedding of market state that then feeds a simpler model.

All three need infrastructure most operations do not have: months of high-resolution recorded data, and the discipline to validate under purge and embargo despite the longer training loop.

The practical recommendation

Start with boosted trees on features you can explain. Establish an honest out-of-sample baseline under purged validation. Only then ask whether a more complex model beats it — measured the same way, with the same costs applied.

In our experience the honest comparison closes the question more often than not, and the time saved goes into the two things that actually improve results: better features and better execution.