Financial time-series models are commonly trained on open-high-low-close-volume summaries. These inputs compress the trading process and leave out information about order submissions, cancellations, queue imbalance, and transaction flow.
MicroTSM studies what changes when sequence models are given richer market information. The research uses Level-2 order-book records and tick-level transactions for more than 5,000 Chinese equities from 2015 to 2025, constructing 137 minute-level market microstructure features for evaluation.
Market microstructure data matters
Across the experiments, market microstructure features contribute substantially more predictive information than conventional OHLCV inputs alone. Removing Level-2 and tick-level features leads to a marked decline in forecasting performance.
The result positions market microstructure data as an important frontier for financial AI: richer observations of how orders and transactions unfold can reveal signals that aggregate price-and-volume bars cannot retain.
Scaling begins to appear in financial time series
Scaling MicroTSM from 30 million to 1.5 billion parameters improves multiple return and risk-adjusted performance measures. The evidence is an early indication that scaling behavior may extend to financial time-series modeling.
The trend is already visible at the tested scale. Extending the study to larger models and broader training regimes will require substantially more compute and is a direction for further research.
A domain-specific architecture outperforms generic baselines
MicroTSM is designed specifically for financial sequence modeling and combines a transformer-decoder architecture with gating and other domain-oriented modeling techniques. In the reported experiments, it outperforms generic Transformer variants and other baselines across return, Sharpe ratio, and turnover.
Out-of-sample backtests cover 2023 to 2025. The study also finds that conventional forecasting losses such as MSE and MAE can be weak proxies for downstream trading performance, reinforcing the need to evaluate financial models against the decisions they ultimately inform.