Risk management for high frequency trading

[dgn2]

@J what is the typical length of a trade?

[Johnny]

@dgn: [20 seconds, 20 minutes]



@qazwsxedc: with no disrespect to you personally, if you can learn some risk management rules from experience (which may or may not be implied by your post) then why can my computer not learn equally good rules from a data set that covers a longer period than your trading experience?



My impression so far from this thread is that high frequency traders either don't know how they manage risk or don't want to say how they manage risk. I suspect it's more the former than the latter, with lots of rules of thumb that are tested on precisely one sample path, i.e. real life. Rules of thumb tested in this way can blow up at any time. More, and better, answers please.

[dgn2]

I agree with Johnny about learning from data. I think alluding to rules of thumb is a bit of cop out. Granted if what you are doing is working for now, then who cares, but I am more interested in why so that I can apply techniques to new situations. I want to be flexible / adaptable. If I don’t know how to develop new rules of thumb as the world changes I am more vulnerable.



I would also be interested in what people do. I will try to take the time to sketch out what I am doing, but because my execution abilities are still somewhat limited this is not what Johnny is looking for. I have found that trading at less than 30min creates a need to incorporate execution mechanics into the backtest because the cost as a proportion of the edge is just too high. Once you account for data mining biases and whatnot you are left with something very tenuous. That said, I do think it is possible to do something based on order book mechanics and I have been working on a set of analytical tools to do this. I might as well introduce them here and have you guys tear them up!

[FDAXHunter]

Of course you can learn from data. You could build a model that accounts for much of the extreme risk with very high confidence (and hence implying very large loss assumptions). You can calibrate to bootstrapped samplings of extreme events. You can simulate liquidity risk and you can adjust for operational failures.



In the end, you're going to have a number that's probably not going to be dissimilar to a conservative rule of thumb. Except that you'll maybe feel great about that model giving you 40:1 leverage intraday. But then you recall that Dave, who used to trade down the road in a style very similar to yours, used 30:1 leverage and still blew up.... so you end up using min(Model, Dave).



The majority of blow ups happen because risk management fails. Losing money because you have no edge in the first place or something truly extraordinary happens that no-one imagined in their worst nightmare is one thing. Losing money because you got a little to cute with your risk management approach is something else.



I realize that all this sounds like it's all free-wheeling cowboys that manage things by gut feeling. Call it what you will, but we are very aware of the catastrophic risk one can face and we'd much rather play things safe and not rely on too many assumptions that are perhaps a bit uncertain. You'd have to build quite a complex model to capture all the risk factors in high-frequency trading, which in the end would still require fudging. So, I'd much rather have a set of simple, easy to understand rules and limitations that I'm comfortable with rather than getting too fancy here. Risk management should be solid, simple and pragmatic.

[rowdyroddypiper]

Speaking as someone who is low frequency (2 octaves lower than Barry White low) it sounds like you need to have the experience to know what data are important for your model and once you have the experience it's kind of like eh, why model it up, I know what I'm doing, or at least what I shouldn't be doing. 



We have this problem in low low low frequency space too, though maybe it's reversed by the lack of data to model; you don't have much to begin with so you end up including meaningless info because it's what you have.  There's just certain measures that people who seem to not blow their thumbs off take but damned if I can think of a way to model it to a form that would be more useful to my decision making.  That's not to say that I think it's bad for someone to want to systematize their knowledge or build up their thinking about a problem from modeling the data, it's just another way to learn it.  Just for the record my trade frequency would be [3 Months, 3 Years] so I could be adding a lot of noise here and apologize in advance.

[Johnny]

I agree with RRP and FDAX: risk management based on experience and pragmatism. I'm certainly not arguing in favour of exclusively quantitative or systematic approaches. However, I haven't yet heard many details of pragmatic risk management for high frequency trading; I'm still interested to hear them.

[qazwsxedc]

Johnny: "why can my computer not learn equally good rules from a data set that covers a longer period than your trading experience?"



Because the response of markets to extraneous events changes over time and because you could not get together a large enough sample of severe enough events to calibrate a model for this purpose.



To give a more constructive answer, there are times around certain scheduled news releases when my personal risk management rules require a position size of zero, based on the observation that what I do has no edge at these times. At most other times I work with similar rules of thumb like the ones FDAXHunter mentioned, except that my leverage is lower than his examples. The concrete figures are largely based on a memorable experience on 7/7/05, divided by a big enough safety factor to make me comfortable.

[Johnny]

I agree with your last paragraph, which is in the same spirit as some previous posts. But I'm interested to hear details of some of these rules of thumb.



I also agree with this phrase: `you could not get together a large enough sample of severe enough events to calibrate a model for this purpose', but think it is still true when applied to a human. Both my model and you as a human would, I think, have the same response, which is to say that no models (either explicit code or human learnt experience) fit the data in the event of a 7/7 or similar, so therefore set position sizes to zero, aka get the phk out of Dodge.



Anyway, apart from getting out of Dodge, I'd like to hear more examples of risk management rules applied to high frequency trading. (EDIT: sorry, qazwsxedc also mentioned a rule about have zero positions around - prior to? - economic data releases. Makes sense).

[cronian]

Is it ever safe to apply more leverage on high frequency than you would be comfortable with on low frequency trades? If so, how much?



Is it ever safe to hope something move very far over a short time frame? Is it safe to use greater leverage intraday than overnight?

[FDAXHunter]

Cronian: Is it safe to use greater leverage intraday than overnight?



In most instruments, yes.



Cronian: If so, how much?



A good rule of thumb is intraday risk limits = 3x overnight risk limits. Just as a rough rule of thumb. Obviously liquidity considerations play a major part here.



To Johnny:



You're trying to prevent getting destroyed in one day. In longer term trading, this is not an issue, because the volatility of instruments is very large compared to your expected profit. So, by looking at volatility as your main source of risk gets you a good portion of the way there (in addition, day to day volatilities are more stable than intra-day volatilities). Overnight gaps are by definition included in daily sampling of volatility.



Intraday strategies tend to have lower profit expectations than long term trades, even when normalized to one day. A long term position might have an expected profit of 4-8 bps per day (10%-20% per year, a reasonable number). Unleveraged of course.



An intra-day strategy might have a profit expectation of 1 bps per day, on an unleveraged basis.

So now, to actually make anything like reasonable money we need to leverage, perhaps by a factor of 6, to get to 15% a year. Of course, our risk-adjusted return is better (well, hopefully).



But, in the extreme event, like a 2001-09-11 (New York), a 2004-03-11 (Madrid), 2005-07-07 (Londinium), a 2001-11-20 (FDAX Phuckup), a 2008-03-28 (FGBL Phuckup), a 2004-05-07 or 2004-04-02 (Bond market insanity).



Intraday moves are usually small... with the occasional outlier, but those outliers will approach, or exceed, the close-to-close vol.



Here's an example: The largest move ever in Bunds was 196 ticks close-to-close. Just as an example, [url=/Show%20Post.aspx?PostIDKey=113058]this (What happened to the Bunds?)[/url] has basically the same size, except that it happened in 10 seconds. So, at the extreme, intraday worst case loses rapidly approach worst case daily loses.... but on an unleveraged basis!

Now you leverage this times some factor n (with n>= 5) and you can see why, just looking at a single example, high-frequency traders will constrain their risks not statistically, but just by saying "okay, we don't want to be wiped out and even if Bunds move 300 ticks against me (hasn't happened, but damn well could, largest Bund range in ticks was 262 over the measly 18 years we've traded that shit), I still want to have, say, 2/3s of my capital", so that gives me a max leverage of 10:1.