Risk management for high frequency trading

[Johnny]

Yes, good. That's a beautiful description of the kind of thing I had in mind. Huge gap risk on usually-liquid instruments with levered positions. The description is very clear if I only trade one instrument, or one very closely related set of instruments.



But how about if I (and my computer) are trading a lot of different instruments? My guess would be that for most of these events, the gap risk only occurs for one instrument or a few instruments. But there must also be some rare events in which gaps occur on lots of instruments simultaneously. (Perhaps 9/11, but I was on holiday at the time so I don't remember. I'll check the data). In this case then presumably very very low leverage numbers would be required, to prevent several minor losses accumulating into one catastrophic loss.

[dgn2]

Thank you FDAX. Very good description...one I agree with and am trying to approach in a similar, but systematic and statistical way. I will try to outline in a post over the weekend how I do this and I would love to get positive and negative feedback. I get a starting point with some model and then use my judgement.

[FDAXHunter]

Johnny: But how about if I (and my computer) are trading a lot of different instruments?



You're going to get some dampening. But not a lot. It depends on how closely they are tied to each other (notice I'm using the term "tied to each other" and not the term "correlated". This is intentional. Correlation means nothing in extreme situations).



As we've been moving in Bund space, let's stick to that for now. Assuming that you are long Bunds and /or long Bobls, you really won't get any diversification effects in the extreme scenarios (They are, after all very closely tied to each other). Should you be be long Bunds and short Bobls you do get some risk reduction. However, now you want to look at the most extreme move in the spread (hence why I mentioned "delta, spread delta and beta" earlier). For high frequency trading in fixed income, there's a limited number of things you can do (basically like 3 or 4 points on the curve) and generally speaking, unless you go across currencies, there's just not much diversification you can get.



In equities, it's quite different. The universe of (realistically) tradable things is much larger and it's quite conceivable to be reasonably balanced (or at least not 100% pointing one way). So the diversification effect is much stronger.

It's also much much less likely that all stocks get hit at the very same moment. The top 10 might, or a sector might, but the top 100 in your universe? Much much less probable.

You'd look at limiting the total beta that you can accumulate.

[Johnny]

Seems sensible. And I suppose as usual conservatism is key. For example, when limiting the total beta, to calculate this using a covariance matrix with (very) high correlation estimates.



One thing I have been thinking about recently (and which prompted this thread, as I wanted to see if someone else would say it unprompted) is the role of human intervention in high frequency trading. 90% of the time when the market is doing the usual `known unknowns' kinds of things, my feeling is that good probabilistic modelling would be enough. However, it's the other times, when we have `unknown unknowns' that a human being is needed to do something reasonable in the face of new and bizarre circumstances.



Usually `something reasonable' would probably be simply to get the phk out of Dodge (see previous discussion). So in my mind's eye I can picture a trader with a lever and a newspaper. Most of the time the trader just reads the newspaper. But when something happens, he pulls back on the lever and the computer reduces risk positions commensurately. I think this isn't a million miles away from the ideas of qazwsxedc earlier in the thread. It's also not that far removed from Nonius' recent adventures jumping off rocks in the pitch blackness of the Siberian night. In this circumstance my trading rule is `don't do it'. Smiley