Trained, Frozen, or Absent: Identifying What a Quantum Layer Contributes to Cross-Sectional Equity Ranking (arXiv (Cornell University))

Summary

• The study evaluates a hybrid quantum temporal convolutional network across 2000 Tokyo Stock Exchange equities using 500 out-of-sample trading days.p.1
• On 500 out-of-sample days, two one-sided tests show the trained-versus-frozen circuit Sharpe ratio difference lies within 0.03 annualised Sharpe (p = 0.0055).p.1
• Evaluating the reference study on 241 out-of-sample days instead of 22 days reduces its claimed performance and renders it no longer statistically significant.p.1
• Testing on 2000 US equities across 625 out-of-sample days confirms that the trained and frozen circuits differ by -0.002 Sharpe with a paired standard error of 0.0073.p.17
• Measurement shot requirements depend heavily on the trained cell state, requiring 8192 shots per security at the median cell to preserve ranking compared to 256 on random inputs.p.19

Key insights

1. The trained quantum layer is statistically equivalent to a frozen circuit initialized with fixed random parameters rather than merely indistinguishable from it.p.1
2. None of the evaluated hybrid or classical model conditions are statistically distinguishable from zero gross returns, with the best Deflated Sharpe Ratio reaching 0.17.p.11
3. Withholding the classical encoder representation from the head shows that the trained and frozen quantum circuits are equivalent within a 0.25 Sharpe margin.p.12
4. Extending the out-of-sample evaluation of the baseline QTCNN model from 22 days to 241 days drops the bootstrap confidence interval to span zero (p = 0.07).p.18
5. A second data re-uploading quantum ansatz family confirms that circuit optimization does not improve performance over a frozen state, bounded at a 0.25 Sharpe margin.p.14
6. Cross-sectional ranking models fail to break even under daily rebalancing transaction costs, with turnover running between 34% and 72% per leg per day.p.16
7. Quantum circuit shot budgets vary by up to two orders of magnitude depending on whether inputs come from trained model encoders or uniform random distributions.p.20