Post-trade analysis is where an execution model is supposed to meet reality. The problem is that “reality” in a slippage report is often a mix of two different things: slippage measured from fills that were actually recorded, and slippage reconstructed from spread and impact proxies because the fills were not available.
These are not the same, and a clean headline number usually hides which is which.
Measured slippage is computed against a benchmark price — arrival, VWAP, or a chosen interval — using fills that the order actually produced. It is the most defensible figure, but it is only as complete as the fill tape. Partial fills, cancelled children, and venue-reported-only fills all leave gaps.
Assumed slippage fills those gaps. A model takes the unfilled portion, applies its impact function, and produces a number that looks like a measurement but is really a model output. There is nothing wrong with this in principle — you cannot measure a fill that did not happen — but it should be labelled as such, and it usually is not.
A readable post-trade report separates the two:
- How much of the order was actually filled and recorded.
- Which benchmark each portion is measured against.
- How much of the reported slippage is reconstructed rather than observed.
- What impact assumption was used for the reconstruction.
A report that does not separate these is asking you to treat a model output as a measurement. That is the single most common overstatement in execution analytics, and it is the one editorial coverage here is most interested in.
Again, this is not a claim that AI post-trade tools are dishonest. It is a reading habit: ask which part of the slippage number you could have watched happen, and which part someone had to invent for you.