We develop Performance-Based Shapley Values (PBSVs), which tell us exactly how each individual predictor contributed to out-of-sample forecasting accuracy. My Python package anatomy implements the necessary algorithms and makes it straightforward to estimate PBSVs of any model or combination of models and any loss (or gain) function.
I show that a simple insight can immensely improve the quality of tick-by-tick trade and quote data of U.S. equities. The Participant Timestamp, available since 2015 in TAQ data, can be used as an active (marketable) order identifier.