I’ve not put all that much focus on trade exit strategies as usually have relied on trading signal to determine when to reverse a position (in addition to various risk related parameters). I have a situation, now, where I need to determine the exit independent of the signal. Thus will need to make an exit decision based on other criteria.

As the saying goes, “cut your losses short and let your profits run”. As simple as this sounds is not necessarily simple to implement. Any given price path will have “retracements”, i.e. price will not move in a monotonic fashion. So how does one set the “trigger” to exit the trade such that can ride over “retracement” noise towards more profit?

**Standard Practice
**Standard practice is to use a variety of indicators to determine when to exit a profitable trade:

- retracement as % of “volatility” (recent trading range)

Using recent trading ranges (a proxy to variance) to scale the amount of retracement can work well in certain conditions. - arbitrary limits: max profit, maximum trade time

We may have a view around maximum time or profit opportunity for a strategy, allowing us to avoid a later drawdown to signal our exit. - various technical indicators

**Some Problems**

When trading intra-day observing the market tick-to-tick, the above rules do not work well. High frequency data has many short-lived high-amplitude spikes (high relative to typical price movements). This is particularly evident in periods of the trading day with lower liquidity.

Whereas recent trading ranges usually provide a reasonable view on the current period’s noise level for medium-low frequency data, the same is not generally true for high frequency.

**Optimal Approach
**An optimal approach to the problem is to use a price path model where we can determine the probability of a retracement swinging back in the direction of profit; Or for that matter, determine the probability of subsequent price action retracing prior to any significant retracement. Such a model cannot possibly be right all of the time, but done successfully will have a significant edge over even odds.

The price path model provides the probability of a price going through a level at time t. Given that can ask:

- what is the probability of the price going through a level within a time period
- what is the probability of the price remaining within a corridor within a time period

The above allow us to make optimal exit decisions based on risk considerations (the corridor) and likely (or unlikely) movement in the positive direction.

Embedded in such a model, though, is an accurate view on the price process within a near window, a view and a strategy in its own right.

I developed a general calibration and prediction framework for the price path model, however, the price process SDE needs more work (although it shows good predictive behavior in certain market conditions, does not handle all well as of yet). Are there better alternatives for the short-term?

**Mean-Reversion Collar
**Short of a semi-predictive probabalistic model the next best thing might be to make use of our recently developed mean-reversion envelope. The envelope can be tuned to various cycle periods and amplitudes.

Noise exhibits as a mean-reverting process around some evolving mean. We can tune the envelope to encompass the level of noise we wish to ignore. The projected mean therefore indicates the overall direction of price on average. We can use this then to determine whether to carry the trade forward through a retracement or not.

All these mean tracking discussions are very interesting (truly, they are). I wonder though – in conditions like those shown, how well do these techniques allow for extrapolation? I mean, obviously they achieve good interpolation, but they are only useful if they have predictive power, and you’ve only touched this point in one previous post. It seems to me that an analysis like the one used here misses the point of capturing the mean reversion unless the initial smoothing one does (regression or filtering or whatever) is magically “correct”. BTW for this kind of smoothing KRLS comes to mind – both fast and smart.

I read the 2003 paper by Engel, Mannor, and Mier on KRLS. Their approach to sparsification is clever and the results are interesting. Looks like a useful algorithm for prediction and addresses some of the issues with SVM.

I could see this as providing a possibly better prediction on the future direction of the mean even for the univariate case, and definitely even more useful with the inclusion of other factors.

As I want to define the mean in terms of the mode of a MR corridor am not directly using a regressor to determine the mean, rather am using the PLSQ regressor to determine minima and maxima a-posteriori and then determine a curve that represents the mode through those points.

I have been thinking about how could still take a similar approach with either a SDE or kernel based regressor — the approach of parameterizing the estimate of the mean in terms of a MR corridor with some mean amplitude.

It seems to me that with the kernel approach I would have to generate a training set based on the current approach (of mode through min/max), which defeats the purpose if we want to arrive at a more robust process for the mean.

More thinking required …

Thanks for the comment and pointer. This blog is really a thought process for me, so ideas will appear here regardless of maturity, so always appreciate feedback.

I agree with you that the above is essentially a regression, in this case some processing on a regressor with a desired degree of “smoothness” to obtain minima and maxima (something I borrowed from the HHT). It provides an a-posteriori view on the noise or MR within some band.

Is it useful in making a judgement about the immediate future, a-priori? For the lower frequencies, I would say, yes, generally. I am using this in some tests now and has done quite well predictavely for exits. For ultra-HF, perhaps not.

Regarding KRLS, I have not used that before. On a quick read, the concept seems similar to SVM. One of the things that has bothered me about SVM is one has to understand one’s data well enough to determine a kernel that will allow the data to disperse appropriately in the new space for regression or classification. I suspect could spend weeks or months of effort determining a good mapping. That does not discount its usefulness. Any pointers appreciated.

I agree that Kernel based regressor/classifiers, and stochastic state systems can probably offer a better picture than a regression based approach. That said, will require more work than the 3 days I’ve put into it 😉

Any updates on this? I would imagine unsupervised learning would perform in this case, right?

I’ve not looked at this particular approach for some time. The use of the approach really depends on your timeframe. For medium – HF, I look at a model i developed to detect momentum (based on orderbook flow). If the holding period for a signal is much longer and needs to ride over noise, then you need to rely on both good signals and/or 1 or more indicators of price direction.

There are many sub-games on using direction-signals within the bigger picture. For example, for medium freq stuff, I separate signal and execution, where execution is done on behalf of a strategy / signal. The execution algo can make use of HF view on direction to try to optimise its behavior in capturing a position over some time period.

Given that the momentum (HF) signals predicts the short-term price direction, how would u incorporate it into the position management. For instance, if a medium term strategy predicts the price is driving up in its timeframe, which translates into a desirable long position, how would you go about managing these positions along with the HF signals. It seems that signals are easier to generate while we still need to insure the optimial inventory that one would like to keep all along.

Well, there are many different schemes, depending on what u r doing. With HF, you are usually running a market making strategy of one sort or another. You can use momentum for the following:

risk signal: adjust your offering such that will not be adversely selected risk signal: flatten out contra-directional inventory before momentum moves the price too far away take an aggressive position

One can also use short-term momentum to direct for better execution.

Hi Tr8dr,

Thanks for the very informative blog. Regarding your momentum strategy, is data from the instrument you trade sufficient or you also need to use data from other instruments/markets?

I imagine that market depth data is huge especially given that there’re many exchanges out there. Do you only analyze data from a single exchange or you need all exchanges to have a good signal?

Thanks,

Vlad

Vlad, I do use information across assets and markets for momentum. To get the most accuracy you want to make use of all sources of information. In more recent regimes finding tradable information on a single asset for short time periods can be difficult without also looking at related assets.