A compact model that reads raw intraday market data ranked Chinese stocks better than six matched rival models in a 2024 test, according to a new preprint. But the result came with a practical warning: when researchers measured the return available after the signal could actually be acted on, STRATA's long-short spread was indistinguishable from zero.
On the study's main test, after scores were adjusted against eight style factors, STRATA reported a rank IC of 0.0728, an IC IR of 1.128, a Signal LS Sharpe of 12.85 and a stress IC IR of 1.030. Rank IC is the daily cross-sectional relationship between the model's ordering of stocks and their next-day returns. IC IR combines the average daily relationship with its variation across days. The figures are means across three training seeds, with the accompanying uncertainty reported as a sample standard deviation.
A model built from raw inputs
STRATA has 244,633 trainable parameters. It takes five trading days of raw five-minute bar and order-book data and produces a next-day cross-sectional stock ranking without hand-crafted features. Its design combines a five-branch stem, four selective state-space blocks and a four-path readout.
The data came from mid-capitalisation Chinese A-shares between 2 January 2019 and 31 December 2024. Across 1,456 trading days, the panel had a median of 995 to 999 names per day, with a minimum of 878, plus 48 bars per stock-day and 25 raw fields per bar. Constituents were selected using point-in-time membership, and the study applied no survivorship filter.
The nominal training window ran from 2019 through 2022 and covered 972 trading days. Validation used 2023, with 242 days and 240,204 samples, while the held-out 2024 test year contained 241 evaluable days. After an embargo-related truncation, the training set contained 955,061 samples.
Before calculating the metrics, researchers removed non-finite scores, capped each day's extreme values at three standard deviations from its cross-sectional mean, standardised the scores and residualised them with ordinary least squares against eight style factors. In practical terms, the reported ranking was evaluated after removing the part associated with those included controls.
STRATA led the matched comparison
The six baselines were kept within 5% of STRATA's parameter count, with realised sizes from 244,317 to 247,873 parameters. All seven model arms used the same data, preprocessing, loss, optimiser, training budget, early-stopping rule and evaluator. That design made the comparison one between complete encoder-and-readout systems, rather than a test of model size alone.
STRATA led on all four reported metrics. Against GRU, its rank IC was 14.9% higher, 0.0728 versus 0.0634; its Signal LS Sharpe was 17.7% higher, 12.85 versus 10.92; and its stress IC IR was 13.4% higher, 1.030 versus 0.909.
Seed-level rank-IC tests gave p-values of 0.014 or lower against every baseline, while paired daily tests reported Newey-West statistics from 4.40 to 6.29 and p-values below 0.001 against every baseline. The uncertainty summary was based on three training seeds.
The model's lead also came with less score variance explained by the eight included price-volume controls. STRATA's explained share was 0.172, compared with 0.180 to 0.220 for the five sequence baselines, alongside a residualised rank IC of about 0.073. The authors say the lead was not explained by greater exposure to those controls, although the check covers only those eight.
The timing changed the interpretation
The central caveat lies in the return convention. The model's score is produced only after the final five-minute bar of the signal day, while the next-day close-to-close label includes an overnight move. That makes the exercise a prediction benchmark, not an exactly replicable trading return.
For STRATA, the overnight component had a rank IC of 0.1275, a long-short spread of plus 75.6 basis points per day and a Sharpe of 24.13. The executable session had a rank IC of 0.0215, a spread of minus 1.6 basis points per day, a Newey-West statistic of minus 0.37 and a Sharpe of minus 0.32. In other words, most of the reported long-short performance was tied to the overnight segment, while the executable spread could not be separated from zero.
Changing the target to the executable session did not change the ordering of the seven architectures: STRATA remained first. Its rank IC was 0.0215 versus 0.0162 for GRU, a 32.7% margin, but its long-short spread remained minus 1.6 basis points per day and was indistinguishable from zero.
In a closer comparison with Mamba, which shared STRATA's selective state-space backbone but used a linear stem, a single decay initialisation and mean-and-last pooling, Mamba's rank IC was 0.0615 versus 0.0728 for STRATA. The reported gap was 18.5%, but the study did not isolate the contribution of each design choice.
What the study does not establish
The authors frame the findings as evidence for a prediction-model comparison, not a deployable strategy. The diagnostics did not include transaction costs, slippage, market impact, price limits or borrowing constraints, and the style analysis covered only eight price-volume controls. The reported close-to-close long-short performance therefore should not be read as a tradable result.
The paper is a preprint identified as arXiv version 4, dated 14 September 2026. On the evidence supplied, STRATA is a stronger predictor under this evaluation protocol, not evidence of a profitable strategy that can be deployed as tested.
Paper data and sources
Original title: A Compact Selective State-Space Model for Cross-Sectional Stock Return Ranking from Raw Intraday Bars
Authors: Mingju Chen, Enze Zhang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text