Fair Spread

Anecdote from a strategy that does not run anymore by a desk that does not exist anymore at a firm that is no more:

When quoting non-tick-bound instruments (i.e., securities with a spread that typically exceeds one tick) one does not only have to have a fair value but also a fair spread. How wide do I have to quote around the fair value. In this story, the baseline for the fair spread was the trailing 20-day average spread within the current 5-min intraday interval: Compute the average (median, actually) spread between say 9:00 and 09:05 every day for the previous trading day for this stock and calculate the average. This is a useful baseline for an appropriate spread for this stock at this time of day.

The strategy made some adjustments to this baseline based on the toxicity of the market. For example, for French stock it would multiply this average by a different factor that for Italian stocks. All this was done by a script which ran every night and generated a daily coefficients file.

The country-specific adjustments were made by a simple pattern match. If the company-internal symbology ended in ":FR" then it was a French stock etc. One summer, the internal symbology was revamped (":FR" became ".F" or something along these lines) and the quant responsible for these scripts left the firm. Yet, all seemed fine.

After a few months we wanted to consolidate a few ops scripts. It turned out that the script kept running for a while after the symbology change but the change had not been reflected in the pattern matching. Essentially, none of the adjustments were done properly. In addition, after the a while the script started to crash every day. TLDR: The strategy was running with a garbage and outdated parameters file.

Remember: the pnl curve did not change.

Anyways, when reminiscing over this story, I wondered how important a model for the spread is and how useful the same trailing 20-day average is. Fig. 1 shows the T+10 s markouts from the market makers' point-of-view for DAX constituents along the left-hand vertical axis. The horizontal axis is the distance of the filled order from the pre-trade mid relative to the historical average half-spread at this time of day. The grey bars show the distribution of the traded notionals.

We see that even this simple fair spread "model" captures a lot of the markout differences. Markouts are significantly worse when the spread is tighter than the historical average at this time of day. Orders far from the touch include cases where the orders were on the BBO but the spread was wider than average but also orders on back-levels which were filled in a multi-level trade. One can probably do a lot better than just the 20-day average over 5-min intervals.

Also noteworthy is that the overall average markout is -0.46 bps even before fees (which can be zero for liquidity providers that meet certain criteria). One could interpret this as liquidity being offered too cheaply.