2/ One of our most important principles is "understand your edge." Your hypotheses should be based on simple priors, or you will just overfit to noise. This is a creative part of quant trading. Building strong intuition about markets can filter out many ideas upfront.

3/ The story here was simple: We make money doing HFT because we help smooth out supply and demand imbalances. Therefore, our pnl could be a less noisy signal on the imbalance of flow in the market, or of retail vs sophisticated flow. An interesting enough idea to measure.

4/ We have an internal dashboard using 5 minute snapshots to graph live pnl. So we were just extra lazy and took this data over half a year as our inputs. Our intuition about what to do: pnl is noisy trade by trade, but when smoothed it can be a longer horizon signal.

5/ This is great because any alpha we find should be orthogonal to our HFT trading. After all it's not very exciting to frontrun your own strategies. The simplest idea is to regress sliding window pnl against future returns. We measured x and y ranges 5 minutes to 2 hours.

6/ There was a noisy effect, barely noticeable. However, focusing on the symbols that did well versus not, there was a clear pattern. The effect was weak to non-existent on high volume liquid majors, but relatively strong on the less liquid alts.

7/ The actual effect was startlingly strong after this simple filtering: effect size tens of basis points per day with strong statistical significance and high capacity! However, this was a weird signal: it was strongest when our pnl was highest, and the beta was negative.

8/ The story kind of makes sense. Our pnl spikes when retail is aping in or panic dumping, because that is the best flow to be providing to. On average when volume spikes retail is buying, and retail price impact reverts to some extent, so you should short when HFT pnl is high.

9/ Putting on this trade is quite tricky though. You want to look at perps that are listed during these crazy times and consider whether the funding eats into your pnl. Markets are pretty efficient overall. Our pnl signal must have correlation with mid frequency traders' signals

10/ Anyway, the actual adjustments to be made are in the context of a portfolio of assets, biasing towards shorting when pnl spikes. There are other interesting effects too: the derivative of pnl, or other similar measures of retail flow.

11/ A big lesson here is the power of private data. If you have a unique insight into markets through your interacting with them, use it. A classic tradfi analog is using iceberg fills on CME to trigger on other venues, since private feeds are consistently faster

12/ Another lesson is to branch off from strategies that work. We focus intently on HFT. How could we possibly compete with the mid frequency experts? This strategy was a concrete way for us to move towards this unfamiliar territory with a crutch from our HFT edge.

@chameleon_jeff Really cool thread. I guess you short when the price went up and long when the price went down? (you only mention shorting)

@Bap_Ldko Thanks. The input is our pnl not the price. Our resulting signal basically never tells you when to long. The asymmetry is interesting

@chameleon_jeff Is it a similar idea like trading the volatility?

@Kevin_ZKP Pnl and volatility are definitely correlated but the pnl contains some information about the flow behind the volatility

@chameleon_jeff When u regress PnL → future returns, did you find derivative-of-PnL (acceleration of flow imbalance) stronger than level itself? Or did it just amplify noise? Sorry for jumpin out of nowhere but i am in finishin my HFT bot still in papertrading mode and i am fine tuning and

@chameleon_jeff Q: HFT in this sense is getting lifted somewhere only if you have probabilistic edge to get the risk back immediately at another exchange/super similar pf of coins on the same exchange? I ask bc the flow imbalance you speak of would make most classic mm strats lossmaking too

@chameleon_jeff @threadreaderapp unroll #HFT