Mid Changes - When and How?

Fig. 1 below shows the distribution of the interval between consecutive mid changes - split by the mechanism of the start-of-interval and the interval-ending mid changes.

There are three ways for the mid to change: (a) a trade trades all volume on the top level (and possibly more on deeper levels), (b) the last (or only) order on the top level is deleted, or (c) a new price-setting order is inserted. These are denoted as "TRD", "DEL", and "ADD", respectively in Fig. 1.

For example, "TRD-ADD" corresponds to the case where first a full-level trade moved the mid followed by an insert of a price-setting order.

Interestingly, each of the nine permutations has characteristic peaks - almost like an emission spectrum. For example, in the light blue (ADD-DEL) histogram we see a peak at 1 ms. This means that 1 ms after a price-setting order is added, it is deleted (strictly speaking, the delete could be on the opposite side). There is no technical reason - orders can be canceled before the ExchangeOrderID has been received in the ACK by using the ClientOrderID. So it seems like the algo(s) behind this have an internal timer essentially implementing a minimum resting time.

Other algos without such a logic cancel their order within less than 10 microseconds (left-most peak of the light blue line). This means that they are canceled long before anyone could have seen them in the public market data and traded against them.

Another interesting peak is that at 100 ms in the dark blue (ADD-TRD) histogram. There seems to be an algo which is configured to wait 100 ms before trading against a new level.

By far most prominent is the sub-1 ms region of the orange (TRD-ADD) histogram. These are orders which try to establish a new level after it has been traded away, i.e., establishing a new best bid at the former best offer price. These scenarios are highly latency competitive, e.g. see here for more details:

Fig. 1 is for the Euro STOXX 50 future (FESX) and other instruments show different patterns reflecting the fact that other algos or participants are at work.

Technical note

TransactTimes (t7 in EIBO terminology) were used. These are influenced by the business of the exchange and can exhibit hundreds of microseconds of variance. As a result, the spectrum for these very short durations is somewhat blurry and not as crisp as the peaks further to the right.

Source of data: A7 Analytics Platform

#microstructure #hft #algotrading #lowlatency #derivativestrading #eurex #quantitativeresearch #marketdata #marketmaking