Satan's Kimchi
Mix the following ingredients:
and the result is what is shown in Fig. 1 below.
Assume you have the largest order on the level - regardless of queue position (pro-rata matching!). Now someone trades against your order's level. Since your order is the largest, it's fill will be disseminated first (20 μs between subsequent fill messages). Based on the fraction of the order that is filled and the total volume on the level, you can infer ahead of time - before anybody else - the total volume of the trade. You can use this information to trade e.g. other maturities or other correlated instruments.
Now suppose your order was the second largest. You would still get the same amount of information but 20 μs later - the latency advantage versus the public data would be worthless because it was fully exploited by the owner of the largest order. There is hence huge value in being the largest order - even if just by a small margin.
And so the inevitable happens:
This cycle is only broken when one of the participants reaches some maximum order size or risk limit. At this point, if one cannot be the largest order anyways, one might as well amend the order way down. But there is now no need for the unchanged order to be this large and so it is amended down to be just a bit larger than the amended-down order. Then a new cycle starts. Fig. 1 only shows a very short interval. But imagine hundreds of these cycles in quick succession.
In the samples I found, there were only ever two participants to this one-upmanship game. Imagine if three or more participants were doing this!
The WSJ article linked below shows how this led to an explosion of the market data volume (also see Fig. 2 below). The CME "solved" this by capping the number of transactions that could be sent (the details are a bit technical but this is what it comes down to). Plus, this one is for the history books - the Eurodollar future (GE) is no more - the last contracts expired end of June this year.
#marketmicrostructure #futures #fx #cmegroup #marketmaking #algotrading #lowlatency
References:
Head?of Quantitative Analytics
1 年Better late than never, here is a chart showing the ridiculous jump in market data updates for the Eurodollar futures.
Head of Quantitative Analytics at Databento
1 年This comes in many variations. The most hilarious is when two silos or strategies in the same company are the two parties involved and/or one of them has a GOTO 1 condition, causing them to breach messaging thresholds in a tight loop.
Core Software Developer/Engineer at Emergent Trading
1 年An interesting part of this is that per CME documentation, if a level is fully filled the fills are disseminated by time priority, not by size. In theory this would mitigate the problem to some degree, but the CME doesn’t follow the rules on GE, SR3, and ZQ products, and always sends fills by order size, even if the level is fully swept.
On-chain liquidity producer
1 年Olivier Guéant, this is a bit further from what we discussed but interesting nevertheless.