|
|
|
February 1, 2010 The last five months have been a blur, but we have seen major break-throughs on several fronts. The reason for this is quite simple. We have done an incredible amount of experimenting. Consider that we have been doing experiments on automated trading systems for more than ten years. Some of the early tests only had one trial, but most of the later had thirty or more. Up to September 2009, we may have run a thousand trials all tolled. We thought that was a lot because it was far more than anyone else we knew and we had been doing it for longer than anyone else, but our horizons have been changed. Since September 2009, we have run approximately twenty thousand trials in approximately one hundred different experiments. Those numbers were simply inconceivable six months ago. This is entirely due to the fact we have reduced both the cost per trial and the time required per trial to an insignificant amount. When the time to run a trial went from hours to less than a second, the cost went to zero, and we entered a brave new world. This has been fabulous, but it has taken some adjustment. Where we once had at least a couple of weeks to explore the results of one experiment before planning another, we now have a couple of hours at most. Where we once had a few dozen trials in a complicated experiment, we now have two hundred trials in a typical experiment. Where we once had to work for weeks to generate data to study, we soon needed to create and use new tools to handle all of the data that appears right in front of us every day. Recent Findings Long Mean Reversion Works Most of our work has been on the family of short term mean reversion systems described in the books of Larry Connors and Cesar Alvarez. In general, we found that the methods chosen by Connors and Alvarez have been astounding robust since 1990, but mediocre at best before the current era. We now believe that the change in commission structures around 1990 added liquidity to the market, and that mean reversion became the dominant behavior because of that. We believe that the data of the last twenty years is more relevant and thus more predictive of the future. Since 1990, short term mean reversion has clearly been a winning bet sixty to ninety percent of the time depending on the effectiveness of the indicators that you choose. This is very nice for an even money bet with max draw-downs of considerably less than ten percent. Amazingly enough, this strategy works using any of the indicators proposed by Connors and Alvarez. In many cases, we could not produce losses even with deliberately pathological parameter settings. The systems consistently make money, but there is a catch. The return for the time invested is quite impressive, but in all of these systems, it is hard to keep your money in the game. We have made great progress in increasing that return (see Ron's recent article in Futures magazine), and we are working at keeping the money active. Short Mean Reversion Is Not as Robust The short versions of the Connors/Alvarez systems also appear to work with the correct parameter settings, but they are much more prone to losses and the losses tend to be significantly larger. We would like to employ the short strategies because they would make it a lot easier to keep the money at work more of the time. Unfortunately, the trade-off does not seem to be worth it. Short strategies are thus not a focus at the present time. Averaging In Doesn't Work We also tested the Connors and Alvarez averaging in system, T.P.S. The short story is that it cannot work under most conditions, and never works under test conditions. Here is the link to the full PDF file for that report: Does_Averaging_In_Work.pdf. It has generated quite a bit of controversy, but the numbers speak for themselves. Long Mean Reversion System Scheduled for Live Testing All of these experiments are leading to something practical, or so we hope. We have a very powerful, low-risk mean reversion system that we hope to launch in live trading this month. Back testing has been very successful and exciting, but we still have a few money management issues to work out. At the moment, we are exceptionally pleased with the exit strategy. The chosen entrance strategy performs very well, but we still think we have some room for improvement there.
October 18, 2009 We released a new study showing the application of Design of Experiments on a simple technical trading system outlined in the book, Short Term Strategies That Work. We used the example from the book so everyone could have the same look at the strategy, and try their own optimization process for comparison purposes. We wanted to demonstrate that we could find parameter settings that improved on the author's results, and that we could do that in very short time. Mission accomplished! If you don't count the time it took to set up the simulator, we found this set of optimal results in a couple of hours using the quick, dirty, and very easy alternative of end-of-day data. This article appeared in the February issue of Futures Magazine, and can be online at http://www.futuresmag.com/Issues/2010/February-2010/Pages/Using-DOE-method-in-trading.aspx . Related Links: |