Skip to content
The Algotrading Book

Chapter 160: SwAV for Algorithmic Trading

Chapter 160: SwAV for Algorithmic Trading

Overview

SwAV (Swapping Assignments between Views) revolutionizes Self-Supervised Learning by introducing Online Clustering. Instead of dragging individual continuous feature vectors directly towards each other (as in SimCLR or MoCo), SwAV trains the network to map features to a set of learnable Prototypes (Cluster Centers).

This approach scales incredibly well and introduces immediate interpretability: the model natively learns “Market Regimes” (e.g., bull trends, high-volatility sideways chops, bear sweeps) via the discrete prototypes.

Core Mechanisms

1. Prototypes (The Codebook)

The model maintains a matrix of weight vectors called Prototypes. Representing structural states of the market, each produced continuous embedding vector is compared against these Prototypes via cosine similarity to find the best match.

2. Sinkhorn-Knopp Algorithm (The Balancer)

If the model was left to its own devices, it would succumb to “representation collapse”—assigning every single stock chart to Prototype #1 (e.g., just guessing “Flat Market” constantly).

To prevent this, SwAV uses the Sinkhorn-Knopp algorithm over the batch. This is an Optimal Transport mathematical technique that enforces an equipartition constraint: it strictly guarantees that charts in a batch are distributed evenly across all available prototypes.

3. Swapped Prediction (The Loss)

For a given financial chart, we create View A and View B using data augmentations (like adding noise or jitter).

  1. We compute continuous features for both: $z_A$ and $z_B$.
  2. We use Sinkhorn-Knopp to assign View A to a hard cluster: $q_A$.
  3. We assign View B to a hard cluster: $q_B$.
  4. The Swap: We train the network such that the continuous features of A ($z_A$) can accurately predict the discrete cluster of B ($q_B$), and $z_B$ predicts $q_A$.

Trading Advantages

  • Native Regime Classification: Your trading bot immediately gets discrete “States” ($1-K$) classifying the current market, rather than raw floating-point gradients that require an external K-Means pass.
  • Batched Efficiency: The Sinkhorn algorithm stabilizes gradients significantly better than point-to-point contrastive losses.
  • Interpretability: You can visually inspect what specific market patterns trigger Prototype 1 versus Prototype 7.

Contents

  • python/model.py: Implementation of the Prototypes, Sinkhorn-Knopp online optimal transport, and the Swapped Cross-Entropy loss.
  • python/train.py: Training loop optimizing the continuous CNN alongside the discrete prototypes.
  • python/evaluate.py: Verification of clustering health and regime assignment overlap.
  • rust/src/: High-performance Rust inference pipeline assigning real-time LOB/Price ticks to the nearest SwAV Prototype instantly.

References

  1. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., & Joulin, A. (2020). Unsupervised Learning of Visual Features by Contrasting Cluster Assignments. arXiv:2006.09882.