TRA: Learning Multiple Stock Trading Patterns with Temporal Routing Adaptor and Optimal Transport

TRA attaches a router and a set of predictors to an existing stock-prediction backbone so that one model can represent several distinct trading patterns instead of averaging them into a single signal, with the assignment of samples to predictors learned as an optimal-transport problem. Published at KDD 2021 (Proceedings of the 27th ACM SIGKDD Conference, pp. 1017–1026, doi:10.1145/3447548.3467358).

The model is a first-class citizen of Microsoft Qlib rather than standalone research code: it lives at qlib/contrib/model/pytorch_tra.py with configs under examples/benchmarks/TRA/ for both the Alpha158 and Alpha360 feature sets, and Qlib’s public benchmark table reports its information coefficient, annualized return and maximum drawdown alongside the platform’s other models. Qlib is MIT licensed and actively maintained, with its most recent commit in July 2026.

Two things to know before treating it as a reproduction target. The README section that reproduces the paper is explicitly labelled “Usage (Not Maintained)” and its result table was generated by qlib==0.7.1; the supported path is the qlib.workflow integration, whose current benchmark numbers are not the paper’s numbers and should not be mixed with them. And three of the four authors are at Microsoft Research, the same organization that maintains Qlib, so the in-Qlib benchmark is not third-party replication — the KDD acceptance is the independent signal.