Temporal Routing and Optimal Transport for Stock Prediction

Summary

The document describes Temporal Routing Adaptor (TRA), a way to extend a stock prediction model so it can learn from different patterns in market data. It notes that momentum and reversal behavior may coexist, which challenges the assumption that observations follow one stable, independent distribution. TRA adds multiple predictors for different patterns and a router that assigns each sample to a predictor.

The router uses representations from the backbone model’s hidden layer and the predictors’ historical forecast errors to infer which pattern applies. An iterative optimal transport procedure assigns samples to predictors under specified proportions, seeking to reduce overall prediction error; those assignments then update the predictors. The document reports that TRA improved prediction performance and investment returns over Attention LSTM and Transformer baselines in stock ranking experiments. It provides no dataset details, numerical results, or discussion of transaction costs and out-of-sample robustness, so the claims cannot be independently assessed from this summary alone.

Key ideas

• Stock data may contain opposing patterns, such as momentum and reversal, that a single predictor may not capture.
• TRA combines multiple predictors with a router that assigns samples according to inferred patterns.
• The router uses backbone representations and historical predictor errors as inputs.
• Optimal transport iteratively assigns samples under proportion constraints to reduce aggregate prediction error.
• The document reports improved results over Attention LSTM and Transformer baselines, but gives no numerical evidence or implementation caveats.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.