Bullets Dodged, Bullets Caught

Suppose you have a resting order in the limit order book. Your fair value model suggests that you should pull the order. But the decision of whether or not to send a cancel request is a bit more subtle.

There are two possibilities: (a) your model is correct and (b) your model is wrong.

The difference is that in the latter case, you cancel will always succeed (and you will have given up queue priority unnecessarily). In the former case, however, when a cancel is indeed justified, there is a fair chance that your cancel will arrive too late at the exchange - that the order has already traded. So you have to evaluate the expectancy of the cancel taking into account the probability that it will have no effect.

Fig. 1 below shows the frequency of timely (blue) and late (grey) cancel requests vs. the gateway-in-timestamp difference w.r.t. the trade which would have or did fill these resting orders. The bottom panel shows the markout PnL (trade price vs. T+10 s mid) associated with each latency bucket. The bars to the left of zero is the markout PnL that the order would have yielded if it had not been canceled. The bars to the right of zero is the actual (but unintended) markout PnL for traded orders whose cancel request came too late.

Observations

1. Timely cancels (blue) bars are spread fairly evenly over the 1 ms interval before a trade event in which they would have been filled. The distribution of late cancel arrival times (grey bars) on the other hand is strongly peaked. I assume that the stream of timely cancels is what triggers the eventual trade in the first place whereas many late cancels are only triggered very late ahead of the trade or by the trade itself.
2. The hypothetical (if they hadn't been canceled) markout PnL of the timely cancels nets out to a red zero. The (actual) markout PnL of the late cancels on the other hand is consistently and strongly negative. This fits the narrative that successful cancels comprise all instances where they were not necessary (model was wrong) but only some of those that would have been. The late cancels comprise only cases where the model was correct and were a successful cancel would have been appropriate.
3. Curiously, successful and late cancel immediately before and after (within a few microseconds) are equally frequent and - counter-intuitively - mark out similarly negatively.