Mean-reversion stop-loss methods

We have also seen this for mean reversion systems. We have tested a large number of stop-loss schemes and failed to find one that improves profitability (expected return increase) in walk-forward testing.

This obviously does not mean that they can't be found. Moreover, a stop-loss will truncate your S/L distribution which may be helpful even if it does not improve the expected return. Or it can be used to avoid disasters (but this is more of an insurance, and as we know insurance cost money on average).

IMO, for mean-reversion systems the problem is inherent in the underlying assumption - you enter a trade when you move away from a mean assuming it will revert back towards the mean. If it moves further away from the mean it's hard to argue, under assumption of mean-reversion, that it will be less likely to start to move towards the mean.

Hugin

 

No, it is not harder to argue that it will be less likely to move towards the mean. You're making the assumption that every draw away from the mean is sampled from some empirical distribution with the same (or similar) variance, but that is most certainly not the case. Since the variance (or volatility) is also a function of time, you cannot say that you drew from some empirical distribution that's parameterized (let's say, in terms of moments about the mean) similarly. This makes sense intuitively, because news changes things substantially. As an example, if the variance at one step into the future is 10x as large as the current variance, there is no argument for it to be more likely to revert.

What we all assume is that the probability distribution at time t+n is roughly the same as the distribution at time t, the point at which we entered the trade, and that repeated draws at each step forward in time are happening on a probably distribution that is /almost/ the same. In most cases, this is true. But in the explosive case I've outlined, the distribution at time t+1, you have no idea what the real distribution is. Condition on the previous time and attempting to forecast, you are still subject to model mis-specification risk and outliers.

 

Actually I agree with this, but it does not give much help in creating a systematic stop-loss rule.

For example news works both ways, i.e. you may gain from them or lose. On average news will probably be close to neutral for your expected return.

Also, if volatility increases dramatically (due to news event), can you really say that it is more likely to continue inreasing than to fall back from current level?

Since such large changes in volatility (or price) are uncommon they provide very few examples to use in order to create a systematic stop-loss rule. And the impact on overall P/L will not be large if the stop-loss rule only affects a very small fraction of trades. We have had a hard time to find a systematic stop-loss rule for mean-reversion trades.

How to react to them also comes down to your position sizing, i.e. how much you risk when the trade has gone against you.

Hugin

 

If volatility rises dramatically, you can say it's more likely to stay volatile because of volatility clustering; whether it's sustained volatility in on direction or another is another issue. I think this is why I notice longer trades: volatility is sustained (each successive draw is farther and farther from the mean.)

I mean, yeah, I acknowledge what you're saying. I just want my cake and be able to eat it too.

What sort of skewness do you see in your pnl results from a mean reversion system? I see a distribution centered above zero, with slight negative skewness [from news], but positive expectation.

 

part time based.
part price based.
part if it doesn't act like it (normally) should.
and sometimes level 2, or other associated reasons.

 

Mainly backtesting. Stops for different strategies are all over the map in terms of risk/reward, but normally no more than 2x the distance from entry 1 to scale 1, if applicable. I never use time as a stop but do use it to evaluate entries. I mainly trade indices.

 

We don't trade volatility (yet) but for our stock trading systems we have seen small negative skew (but this is actually varying over time - negative for 2006 and 2008, positive for 2007, 2009) and quite large excess kurtosis (~3). The large kurtosis is why we put in quite some effort to try to truncate the down side of the distribution. But no luck so far...

Hugin

 

For perspective, what %'age of your trades land in the green?

 

For our oldest running stock trading system, a long only mean reversion system that’s been running from 2006 and forward, we have a hit-rate of 61% over ~2100 trades. We define a hit as making excess market profit (since this makes comparisons easier between different trends and fits the way we hedge). The set-up has changed over the years but this is mostly without stop-loss and with time based exits.

Last year we created a new system where we tried to get a balance between profit target exits and stop-loss exits to retain a somewhat symmetrical distribution. Profit targets were easy (and increases the hit-rate) but we hit the wall on the stop-loss. So now we’re trying to figure out how to move forward. We’re a bit hesitant over a system where the returns distribution is truncated on the upside and open on the downside (affects hedging and position sizing).

Hugin

 

you could hedge with other baskets. you can test when it's optimal to get out of the basket. you could model and assume price shocks as alternative subset play. you could hedge volatility. you could model longer or shorter plays to also hedge. you could also better select markets to trade.

sounds like you should at least incorporate volume in your model.