I’ve been writing a bit lately about how our own YK Strategy, despite doing very well in real-time in all different market types, fell down badly when the market made an extreme divergence from history.
As I discussed in Filtering Abnormal Markets, the strategy needed a means to measure when the market has moved beyond normal ranges, because a strategy based on historical norms might not be suited for that very abnormal moment. My initial solution was to move to cash when the market moved outside extreme volatility bands (similar to Bollinger bands).
But from the moment I wrote that, I’ve been unhappy (and sleepless) with that solution because it was a light switch. By a light switch I mean it was a filter that was either off or on with no in between. Philosophically I have a real problem with that – I think a big reason that most traditional trading strategies fail is that they don’t see shades of grey.
For example, let’s say we have a strategy that buys when moving average A crosses over moving average B. Well what if A is just a hair above B? Why is that so much different than a hair below B that it justifies taking a position? When we design systems that make such hard black and white distinctions, the likelihood that they’ll work as advertised in the future is greatly reduced.
A NEW APPROACH
My new approach for YK is based on the idea that the majority of the time, the markets are in a “normal” mode when YK will perform very well, but sometimes, the markets become so “stretched” that predictive tools based on norms are no longer valid. There are different levels of stretched. There is so stretched that the strategy should stay completely out of the market, and then there is stretched enough that the strategy should be cautious but still take a reduced position.
But how do we gauge this mathematically?
The chart above shows the S&P 500 in 2008 along with normal (green, +/- 1.5SD) and extreme volatility bands (red, +/- 3.0SD).
The new YK approach says that (a) when the market is within the normal bands, take the full position, (b) when the market is outside of the extreme bands, move entirely to cash (these type of markets are just too unpredictable), and (c) when the market is between the two bands, take a fraction of the original position (0-100%) based on how deep the market is into extreme territory.
Historically, from 2000 to date, this approach would have taken a full/normal position 76% of the time, greater than half the original position 19%, less than half 4%, and moved entirely to cash 1% of the time. It would have had no impact on long-term returns and reduced average and peak drawdowns, while still ensuring (like our original light-switch approach) that the strategy is completely out of the market during times of extreme market stress (a’la October 2008).
How would it have performed this year? See chart below of original (blue) and revised (red) strategies relative to the S&P 500 (green).
I’m sharing this approach because I think it has applications beyond YK. If you are willing to accept the belief that a market in stress is a market that is less predictable, then this approach is a solution that still respects the market’s shades of grey.
Happy Trading,
ms
To stay up to date with what’s happening at the MarketSci Blog, we recommend subscribing to our RSS Feed or Email Feed.
Filed under: Evolving Markets, Trading Strategies | 11 Comments
Much better solution than a binary one. Question: would you be able to base your position size on market volatility, using ATR, Imp Vol, etc? That too seems like a good way to reduce risk.
What measure for volatility are you showing on the chart above? Is this simply the std dev of the S&P closing prices — or something like the std dev of the average true range? Also, what look back period (if any) are you using — or would suggest to use? Thanks for your very informative blog, it has given my many things to think since I discovered it a few months ago.
dan
RE to Steven Place: we toyed with idea of just using pure volatility as a gauge of abnormal markets, but found that the strategy didn’t necessarily have difficulty with pure volatility (absent directionality), but it could have trouble with directional volatility (ie the market meltdown/up). Having said that, something based purely on vol might be more appropriate for a strategy that was (for example) only buying extreme weakness b/c our approach (which implicitly runs away from extreme weakness) wouldn’t work. Thanks for the comment. michael
RE to Dan: it’s very similar to a Bollinger band. I intentionally didn’t give the specific parameters we’re using because I want folks to get the concept, not just copy what we’ve done. I think the right answer depends on the time-frame the strategy is trading in (hold times, lookback periods of indicators used, etc.) Thanks for the kind words re: the blog. If I’ve sparked some new ideas then I’ve done my job. michael
Using a simple Bollnger Band of 60 days and ±3 does give a good answer, but ANY sub-band only detracts from the overall return and does not improve drawdown. So even though it sounds like a good idea, a quick test does not show it makes any real improvement. It does look like we go from light to dark with no gray.
Buy the way, a 60 day BB with ±3 does keep you safe in 1987!!
and you can logicaly say anything over 3 std dev is not a curve fit, but a true measure of out of the ordinary situations.
Thanks
Monty
RE to Monty (first comment): I have no idea how you can draw that conclusion without knowing the trades we’re placing (long/short, allocation). Remember, the point is not to use the bands to trade, the point is to use the bands to know when not to trade, or to trade less than a full position. Sometimes those positions could be short and sometimes long, but the band concept keeps us out (or partially out) of all of them.
I’m not sure I understand the “we go from light to dark” comment either – everything between the “normal” and “extreme” volatility bands is a percentage between 0 and 100% (and everything in between) of the original position.
Don’t get it.
michael
Michael, good to see the position sizing adjusted with this, as I like that for one option of not turning off the light switch.
After your first post, I immediately thought that two bands were needed, to guide position sizing. I did not begin testing it on my own systems, but I will.
Anyway, good stuff.
Side note: if you go to my covestor site and look at the equity curve, you’ll note it is similar to yours, and you will be able to see clearly when the light switch turned off. It has not turned back on yet, for the two systems I’m trading in that account. I think our method for turning off is too harsh though, and am looking forward to messing around with some bands.
Michael,
I understand what you are using the modified Bband for. I just applied the concept to a mean reversion system I had. I removed my normal “trouble” detector and replaced it with a simple BBand of (60,3). I then ran an optimization to sweep the std dev from 0 to 5 in increments of 0.1. I then plotted out the equity curve. What I saw on my system was a very straight line of increasing return right up to the 3 std dev, and then it fell off. If we were to have a gray zone, the equity curve should have had a kink in it where you run into lower returns before they go negative. I did not see that so I am just giving you one point in the universe of a system that had no gray zone. It would be easy to run your system the same way and show the gray zone which I suppose on your system would start about 1.5 and continue to 3 before a drop. One reason I suspect there is no gray zone is because when you look at the returns from a mean reversion system, you usually get higher % gains the farther you are from the MA trigger. So to cut your position in the zone of 1.5 to 3 might not be improving your system.
Thanks
Monty
RE to Monty: That makes more sense…sorry, i get so many crank comments on this blog (most of which never see the light of day) that I assumed this one was out of left field. Again, my sincerest apologies. [p.s. i like people who actually test things]
I’m a little confused on your most recent comment (I’m slow so bear with me), but I have a couple of immediate responses. First, YK is not a mean-reversion system. Let me rephrase that – at this very moment it’s mean reverting, but wasn’t nearly as much just a couple of years ago (it’s moved from very momentum based, to very contrarian…part of the “evolution concept”). So you’re right, a pure mean-reversion program is going to react differently to this idea. I need a solution that fits with the “evolutionary approach” and works regardless of whether the strategy is contrarian or momentum based.
As a side note – interestingly the strategy did just fine in the 1987 drawdown without the filter b/c it was very neutral to that market condition (not long or short)…again, it evolves, but in that case it was probably more luck than anything else.
Second, the part that confused me was how you were analyzing the inner bands. In my proposal (which I probably did a poor job of explaining in the post), within the inner bands, the strategy takes its normal position (which again, could be long or short). Outside the outer bands, it takes no position. Within those bands it scales – a little outside the inner band would be almost a full position – almost touching the outer band would be almost no position. Is that the same way you’re looking at it?
Thanks again for the follow up.
michael
I think I am getting closer to an understanding. Yes, I also switch between momentum and mean reversion. Since 2003 it has been strong mean revert. I did a simplification of your scaling and went in 100% when inside the 1.5% band, and then cut to 50% when between 1.5% and 3% , and then 0% when over 3%. I will set up a smooth scale based on where it is when between them and give it another go. This is very interesting.
Thanks
Monty