I had a system (DF20 reversal) which traded 1 very specific type of reversal. Its trade frequency declined by 50% in the 2nd half of 2012, going from 10% of the total number of trend-changes, to about 5%. I decided to look at the other side of the equation, those 90% now 95% of all trend-changes. Wrote a small piece of software around that system, to capture a bunch of things at each tend-change. Then analyzed that stuff (about 12000 trend-changes in 6 years, Excel is great for that amount of datapoints). I didn't find the 24 patterns at once, of course. Anyway, the key question, that I have been trying to answer in that statistical analysis, is : "At any given trend-change, given its "context", should I go with the trend-change, or against-it". Each pattern corresponds to a specific set of conditions for "context", and the associated statistically most profitable trade decision (go with, or against, the new trend). The system simply takes a position according to that trade decision, and keeps it until a new trade decision leads to reversing the position. I have not found a single "trade management" rule (stop placement) that yields better results, than having no stop, and simply waiting for the "proper time" to reverse that position.
Results for the week ending August, 23rd: - 6 wins ; 11 losses ; net -1855 Unrealized P&L on the open position is about +1450 at this time. Not a good week, Monday & Tuesday were particularly difficult for this system. I understand clearly why (in terms of PA) the system is in drawdown since the beginning of July, this is likely to continue until after the Labor-day week-end, for the time being I have nothing else to do than sit tight. I am growing suspicious of having over-fit the system to its in-sample data, the out-of-sample performance on 1H13 ranges from 64% to 75% of the backtest period (depending on the configuration traded). I have started to explore a different way of using the patterns to make the trading decisions (from fixed-priority to adding individual pattern scores), which seems to show a more consistent performance level on 1H13 vs 2007..2012 (at the expense of ~10% less performance on the backtest period).
Results for the week ending August, 30th: - 6 wins / 8 losses / 3 BEs : net -2625 The week had a good start (+3300 by Tuesday afternoon), then the system got caught short in the overnight spike-up (Tuesday -> Wednesday), reversed shortly after the top, and got another big loss right away (-5290 in 2 trades). On top of that, I had made & switched to a new version last week-end, not realizing that I had made an error in that version, which cost me another 1500 in the form of a -800 loss instead of a 700 win. Even with that error, the performance was marginally better (by $100) than the previous version. I have to admit I felt devastated Wednesday morning. I spent a couple days re-visiting stop strategies, with the same conclusion: over 2007..2012, using stops decreases P&L & increases max drawdown, even though there are periods of time where using stops would locally yield better results (and 2013 falls into that category).
I tried both simple-stop & stop-and-reverse ... stop-and-reverse was even worse than plain simple-stop. I suppose the explanation for it is, that whenever I have spotted a pattern with an edge, the system uses it to reverse. I am left to using stops which placement have no edge, whatever I tried so far.
From start of live trading (March-19) to July 3rd (P&L peak), my average net per trade (live account) was about 70% of the backtest value excluding 2008, on ~200 trades. Since July 4th (last 150 trades), my account has been in drawdown, most of the time less than 50% of the max historical DD - it is only this week that the DD has reached about 75% of the max historical DD (again, not counting 2008). WRT comparing live trades with backtest trades at the end of each week, the live trades match 100% the backtest trades, with on average 1-tick positive slippage vs the backtest. Since I went live I made a 2 significant "performance improvement" upgrades to the system, plus a few minor ones. Although I have not used 2013 data to identify additional patterns, I used whatever 2013 data was available at that time to validate/reject any candidate change. The first significant performance improvement went live on June 1st, and its results on the last 2 months are a little better than the prior versions. The 2nd significant performance change went live last week-end, and although it essentially removes (or reduces) one potential over-fit (going from fixed priorities amongst patterns to a scoring mechanism that takes into consideration all matching patterns, which weights are directly coming from the 2007..2012 statistical analysis), a key factor in the decision to use it was its vastly improved performance in 2013 ytd (despite a 5% reduction in the 2007..2012 performance), for this reason I don't want to comment on its performance up until now. For system-stop planning & drawdown monitoring, I use MC simulations based on a trade-distribution at 1/2 expectancy of the backtest. In that light, my account current drawdown is 1 stdev less than the expected average yearly max DD - in other words, there is a lot of room for more "normal" drawdown.
If you still have no grounds that might call into question your original analysis, and if the max DDs are still less than historical maxima, and if your position sizing is such that you can still hang in there if DD gets even worse ... then just keep trading the system and wait for the good times to return.
Hi Dom question about your always-in strategy. I think thats kind of a cool concept, but it seems like if you only identify certain patterns the program should be involved in a trade only after those patterns are seen, right? So how does the program make decisions if those patterns do not current exist?
The system doesn't make any decision when there is no matching pattern (ie., it just keeps the current position as is).