Trades > Cancels > Inserts
Trades by far have the most signal value; someone spent real money to cross the spread plus exchange fees. Passive order inserts, on the other hand, carry much less predictive value; they are, what economists would call, a cheap signal. After all, the likelihood of getting traded against before having the opportunity to cancel the order are usually very slim (the exception being price setting orders or orders joining a very short queue early in its life).
Cancels are an interesting intermediate case. Naively one would think that cancels must be even cheaper than inserts since they avoid incurring costs. This neglects a resting order's optionality value which depends on its position in the queue. Suppose you cancel an order and insert another one at the same price. The new order would join the back of the queue, i.e., you would have lost queue position. And as I have written previously (e.g. here), queue position has a huge impact on the execution quality (likelihood of getting filled and markout in case of a fill).
This position in the queue was "bought" by a lot of investment in tech (see e.g. here) or patience (see e.g. here). So, cancelling an order is not free unless the order already is last in the queue.
And because the loss of optionality depends on the queue position, one would assume that participants are more hesitant to pull orders the better their queue position is. Or, turning this argument around, if one observes an order being cancelled, then the signal value of this action should increase the closer to the front of the queue the order was. Unlike trades, it is not binary.
Exacerbating this could be that the participants' sophistication increases as one moves towards the front of the queue - think HFT, FPGAs. So, their decisions are likely to be well-informed.
Fig. 1 below demonstrates just that for the Euro STOXX 50 future (FESX) and the Bund future (FGBL) on Eurex. Using A7, we extract all cancels of orders while on the top level together with their pre-cancel queue position and whether or not the level ends up surviving or collapsing. We then compute the probability of the level surviving vs. the volume behind of the order before it was pulled.
领英推荐
Observations:
High-Frequency Market Making
1 年How are survival and collapse defined? Over a certain future time period?