25% of my entire risk capital is allocated to futures trading. Perhaps 15% of that 25% is trend following. Then I also have some exposure to trend following in my long only portfolio as I use it for asset allocation. So 25% is roughly the right figure. GAT
There are sort of two questions here, how come we don't hit our target values precisely, and how often should we revisit them? To answer the first question, unless the distribution of forecasts is identical across time you're never going to hit it out of sample. To get an idea of the noise I just plotted the estimated forecast scalar for mac64_256 as an experiment, and it varied over time between 2.0 and 2.3; over the last 20 years or so it's shown a consistent up trend. The differences here are small enough that I probably wouldn't worry about it. (there is also a subtle issue here in that I am pretty sure the targeting function weights markets with more data more highly, whereas you are taking an average across all instruments equally - this will be especially problematic for slower momentum and carry; a market with better trends or higher carry like Eurodollar will have a higher average forecast, and there is also different time periods of data for different markets.). How often should you revisit them? It partly depends on how much data you use. We have the usual problem here that we want enough data to fit robustly but not so much history that we miss systematic changes. But there is no good reason to expect that forecast values will change substantially every few years. I use all the data for forecast scalar estimation, which means that re-estimating them every year will hardly change the values very much at all. I also cross validate my forecast scalar estimates with randomly generated data, which of course won't be affected at all by extra data. So in practice I tend to keep forecast scalars fixed once I've found values that are roughly correct. GAT
Hello Robert, In one of your blog articles you have a link to your presentation called "The Myth of the perfect trading system". The link was https://www.mta.org/video/the-myth-of-the-perfect-trading-system/ and currently at https://cmtassociation.org/video/the-myth-of-the-perfect-trading-system/ however it seems to be behind a (quite expensive) paywall. Is there a way to see this video for free?
I read this very good post by GAT with interest. Fwiw, I have tried playing with GARCH(1, 1), which was mentioned in the post, and had a similar (if not identical outcome). The vol forecasting (using full out of sample testing) improved (I used equally weighted vol 25 days forward as the test parameter). However, when I plugged the GARCH vol into the full backtest of the system, the risk adjusted returns not only did not improve but actually got worse relative to the simpler vol scheme I was previously using. It was a few days of intensive work, and when I saw the improvement in forecasting, I must admit I became rather optimistic. But ultimately, I would class it as a 'failure' too. At least with respect to improving the trading system returns. It would be interesting to hear if anyone else has toyed with different vol schemes either GARCH or otherwise, perhaps with more success.
To be clear, I wouldn't test something like this based on just the risk adjusted returns. The effect it has on the distribution of returns, skew and kurtosis, and what band the risk outcomes are could be equally important. But yet, as a route for improving returns, improving vol forecasting is probably a waste of time. GAT
Why not use a higher frequency forecasting method like HAR? Most people I know that have tried it found it to be an improvement over GARCH
A fun experiment is to replace the vol estimator with a function that has perfect foresight of future vol. Then see if this improves strategy returns. If not, then it's a waste of time to try. GAT
Yeah, mostly my comment was that it’s easier to shift to a more information rich model than to improve something like GARCH. FWIW, I actually do a lot of “perfect forecaster” experiments where you take some non-price variable, assume that you can perfectly predict it at some horizon and trade on that information. If that yields really good results, you try to actually build a model to forecast that variable. The idea is that it’s easier to forecast something like economic numbers based on some trailing data with high degree of confidence.
So rough 'n ready results where perfect foresight is simply a shift function on ewm volframe. This suggests that vol forecasting may be not hopeless after all. SRs are lookbacks from today. My interpretation is there must be a forecasting sweet spot or local maximum for a given portfolio based on avg hold period?