Push-response anomalies in high-frequency S&P 500 price series

Battle Count: 57

Highly relevant to quantitative trading. The paper identifies specific lag-magnitude regions where intraday price responses are conditionally predictable, directly applicable to short-horizon trading strategies. The push-response surface can serve as both a predictive signal and a risk attribution tool. The findings on asymmetric liquidity replenishment (stronger positive responses after negative pushes) and the identification of momentum vs. reversal regimes at different lags are directly actionable. The paper explicitly discusses trading implications including long vs. short implementation, transaction cost considerations, and the construction of lag-aware, variance-adjusted trade filters. The methodology is designed for the most liquid instrument (SPY), making it immediately relevant to high-frequency and mid-frequency trading desks.

The core methodology is conceptually straightforward: compute price increments at various lags, standardize, bin, and compute conditional means. However, practical implementation requires: (1) access to proprietary NBBO/TAQ data (~2.6 billion events), (2) efficient event-time processing across 7 exchanges, (3) careful winsorization and outlier removal, (4) construction of 3D surfaces across hundreds of lags and 320 bins, (5) symmetric/antisymmetric decomposition with bootstrap confidence intervals, and (6) handling of same-session constraints. The computational infrastructure for processing billions of ticks is non-trivial, though the statistical methods themselves are elementary (binning, averaging, bootstrap).

The methodology is fully described with explicit equations (Eqs. 2.1-3.22), grid parameters, and data processing steps. However, the NBBO data from TAQ is proprietary and not publicly available. The paper does not provide a GitHub repository or code. The exact list of lags, binning procedure, and winsorization thresholds are specified, enabling replication if data access is obtained. The analysis is nonparametric and does not require model fitting, which aids reproducibility. No random seeds or software versions are mentioned.