Trade-level regression analysis of predictive signals in pairs trading

This paper examines the viability of high-frequency pairs trading in China's A-share market using minute-level data across three GICS sectors. Applying a Bollinger-band framework to 4,500 cointegrated stock pairs, we document consistentthough moderategross profitability. We then conduct trade-level regressions to isolate drivers of profitability and identify three key determinants: spread volatility, market-cap disparity within pairs, and multi-frequency return correlations. Further diagnostics show that execution latency significantly erodes returns, convergence reliability closely aligns with estimated half-lives, and exit effectiveness varies asymmetrically across long and short trades. Sector-level results reveal that Consumer Discretionary delivers the strongest risk-adjusted performance, while Retailing features the fastest mean-reversion and highest trade frequency. Robustness checksincluding random-pair bootstraps, threshold sensitivity tests, and stress-period performance during the COVID-19 shockconfirm that profitability is not driven by data-snooping or microstructure artifacts. Overall, the findings provide new evidence that high-frequency statistical arbitrage remains feasible in an emerging market setting, while highlighting the critical roles of execution speed, volatility conditions, and behavioral inattention in shaping trade-level outcomes.