Skip to content
The Algotrading Book

Temporal Contrastive Learning for Stocks

Temporal Contrastive Learning for Stocks

This repository implements Temporal Contrastive Learning (TCL) for self-supervised feature extraction on financial time series data. Instead of generating artificial permutations or noise to create “positive” pairs (like SimCLR or BYOL), TCL leverages the natural arrow of time.

Core Concept

In financial markets, the underlying “regime” or state changes slowly relative to minor tick-by-tick noise. Therefore, a time window $x_t$ and its immediate temporal neighbor $x_{t+k}$ (where $k$ is small) share the same underlying latent market state.

TCL uses this structural property:

  • Positive Pair: $(x_t, x_{t+k})$ — Architecturally treated as different views of the same underlying regime.
  • Negative Pairs: $(x_t, x_{rand})$ — Time windows randomly sampled from distant points in the dataset, representing different regimes.

The network is trained using an InfoNCE (Normalized Temperature-scaled Cross Entropy) loss, but instead of the positives being artificially augmented images, they are naturally occurring adjacent time slices.

Trading Advantages

  • No Artificial Distortion: Financial data is continuous. Artificially injecting random noise or reversing signals to create augmentations can destroy crucial financial meaning. TCL avoids this by using the raw, unaltered data sequence.
  • Temporal Smoothness: Embeddings learned via TCL evolve smoothly in the latent space over time. This makes them exceptionally useful for Hidden Markov Models (HMMs) or regime-switching classifiers, as the model naturally tracks macro-state evolution.
  • Trend Invariant: By contrasting local adjacency with global distances, the model intrinsically learns what makes a specific local trend unique compared to the rest of the historical dataset.

Project Structure

  • python/: PyTorch implementation of the CNN1DEncoder, the target projection head, the InfoNCE-based TemporalContrastiveLoss, and a custom training loop that samples adjacent temporal windows.
  • rust/: High-performance Rust library for real-time temporal feature extraction in production trading systems.
  • docs/: Theoretical deep dive and implementation details.

References

  • Concept inspired by Contrastive Predictive Coding (CPC) and modern Temporal Contrastive Learning frameworks for video and continuous streams.