Does a quantum layer improve stock rankings—or just sit in the model?

A direct test of the quantum layer

Hybrid quantum-classical models combine a conventional neural network with a quantum circuit. If such a system ranks stocks well, the circuit might have learned a useful transformation. Or it might simply provide a fixed nonlinear feature, while the classical parts do the important work. A clean comparison needs to change the circuit while keeping the surrounding model as similar as possible.

Chi-Sheng Chen and colleagues built that comparison into a study of daily stock ranking. Their primary dataset covers 2,000 Tokyo Stock Exchange securities. A temporal convolutional network processes 20 days of 21 technical features, then supplies a compact representation to a quantum convolutional layer and a prediction head. The task is to rank securities across each day, not forecast an exact price: the training labels mark the 200 highest and 200 lowest forward returns in each cross-section.

The key comparison is between a circuit whose parameters are trained and the same circuit held at its random initialisation. The authors also test a parameter-matched classical surrogate, remove the circuit entirely, and run versions in which the circuit is the head’s only input. Five conventional machine-learning models provide broader benchmarks. The conditions share a code path, training procedure and data splits, helping isolate which component changes the result.

What the comparison can establish

The main evaluation uses four rolling walk-forward folds and five random seeds, yielding 500 out-of-sample trading days. On this Japanese-market track, the trained and frozen circuits differ by only 0.008 in annualised Sharpe ratio. A two one-sided equivalence test places that difference inside a 0.03 margin (p = 0.0055). In plain terms, the evidence supports a small difference between these two conditions; it is stronger than merely failing to find a statistically significant difference. The authors disclose that they chose this margin after seeing the walk-forward results, so it is not a fully pre-registered threshold.

The result is not a demonstration that the circuit itself has no effect. Against the model with no circuit, the experiment can resolve differences only within a much wider 0.50-Sharpe margin. The trained circuit, frozen circuit, matched classical surrogate and no-circuit model are not cleanly separated by the study’s tests. Circuit-only comparisons likewise find the trained and frozen versions equivalent within 0.25, but that looser margin reflects lower precision. A second circuit design, which re-uploads data and reads out all eight wires, also finds trained and frozen conditions equivalent within 0.25, but not the narrower 0.03 margin. Its point estimate favors the frozen circuit, the reverse of the primary circuit’s small trained-favoring difference.

None of the eleven main conditions on the Tokyo walk-forward track is distinguishable from zero by the study’s tests; the best Deflated Sharpe Ratio is 0.17, below the authors’ stated 0.95 threshold for a claim. Large differences among random seeds and market periods further complicate model comparisons. The 135-day holdout is more profitable for all tested conditions, but the no-circuit model ranks above the two circuit versions, with no significant separation among hybrid conditions. This reinforces the ablation finding, not a general claim of profitable quantum investing.

A second market, and limits

The authors repeat much of the analysis on a US-equity panel of 2,000 stocks. There, every condition loses money during the walk-forward period, yet the trained and frozen circuits remain nearly identical. The authors also re-score their own earlier model using 241 out-of-sample days rather than its reported 22; its point estimate remains high, but the longer-window uncertainty no longer supports the earlier performance claim. They identify several issues in the earlier pipeline, including split-adjustment direction and missing annualisation in released code.

This is a simulation study, not a demonstration on quantum hardware. Both stock universes have selection limitations: the Japanese competition universe was fixed using a later snapshot, and the US sample begins from a current symbol directory that omits some delisted firms. Those biases are shared across conditions, which helps comparisons but limits what absolute performance says about the broader market. The analysis uses a narrow set of technical indicators, so it cannot rule out a circuit helping with other kinds of financial information or tasks. The authors also disclose that some decisions and added analyses followed initial results, and that an earlier holdout run could not be reproduced; reported results use the reproducible rerun.

The study’s narrow conclusion is that, for this architecture, feature set and ranking task, training the quantum layer did not improve on freezing it. The weaker evidence about removing the layer does not establish that quantum components are universally useless. The paper also estimates that preserving score rankings with finite circuit measurements may require thousands of shots per security at the median trained model, emphasizing that hardware cost depends on the trained model’s actual inputs. The work is a preprint and should be read as a detailed, self-critical benchmark rather than a settled verdict on quantum machine learning.