Neural Networks for Market Data: When to Use MLPs, CNNs & LSTMs
We are selective with deep learning. Expensive to train, easy to overfit, harder to debug. But they shine with rich feature sets and nonlinear temporal patterns.
1. Why We Are Selective with Deep Learning
Deep learning earns headlines, but on market data it carries real costs that a simpler model often avoids. Three in particular shape when we reach for it:
- Expensive to train: requires significant compute resources
- Easy to overfit on noisy financial signals
- Harder to debug and explain than trees or linear models
2. Where They Shine
- Rich feature sets (order book, microstructure)
- Nonlinear temporal patterns
- Large cross-asset datasets with shared structure
3. Types We Consider
MLP: simple feed-forward for tabular embeddings.
1D CNN / TCN: convolutions over time; good at local patterns.
LSTM / GRU: sequence models that retain state.
4. volarixs Stance
Neural Networks are one of the model classes you can select in an experiment, alongside regression, trees and boosted models, and time-series. We treat them as a second tier: fewer configurations, reached for on selected universes or factors once a simpler baseline has set the bar.
The discipline is to never run a deep model on its own. Each experiment is fit on the same walk-forward windows as its simpler counterparts and records train and test R² side by side, so a neural network has to clear a baseline on held-out data before it earns a place. Because regime context travels with every run, you can also ask the questions that matter most for deep models:
- Does the out-of-sample fit actually beat a linear or boosted baseline?
- Does it hold up across regimes, or only in the one it was trained in?