Yes, of course, but the eternal hope of all financial alchemists is that the perfect combination of slightly different tools may one day turn our charts into gold.
Well said! All possible combinations of every conceivable indicator and system theory applications, like identification, adaptation, etc., have already been exploited by quants, especially in Eastern Europe and Russia, where people had the time to play with tools but not the money. The market has defeated almost all those combinations, not allowing any of them to have positive expectancy for more than one year, in most cases rapidly giving back all profits from previous period. The only thing that is left to trade is good, old, plain vanilla price action. This is what HFT does in low latency mode but for longer timeframes it is very hard to find an edge due to the lack of models. HFT model is easy but models, for example for swing trading, based on price action only do not exist or are being kept secret.
If that's the case, then the issue is not so much curve fitting (i.e., fitting to a pattern that is not really there), but rather is one of dealing with a finite window of opportunity. In that sense, trading is like any other business, except that the window of opportunity is a lot shorter than in others. In other words, the problem of strategy design becomes not of curve-fitting (chasing mirages) but rather of non-stationarity -- the patterns are there, are statistically significant, are mechanism based, but they disappear as soon as you start exploiting them. In this scenario which is also well-known and discussed on this board (i.e., non-stationarity), then all one needs to do is find an edge, keep it a secret, but estimate as best as you can the expiration date or window of opportunity. Have several new strategies in the pipeline, and as soon as one expires (even before your PL curve starts going down) you swap with the new strategy and continue the cycle. So instead of coming up with an adaptive strategy that continuously fits just the parameters, you come up with a pipeline of strategies that are proven to work in the last 1 year or so, run it until your estimated expiration date, then swap to the next in the pipeline and keep going. Another approach is to come up with a strategy that is highly non-obvious but works. The expiration date on those will likely be longer so you may be able to make more money on those before moving on to the next one.
Sharpe ratio has flaws, but we found that the Sharpe ratio, together with the Ulcer index, are from all performance gauges we've tried the best correlated to the robustness of a strategy. But normally you have to look at many parameters - one alone is not enough. By the way, I'm surprised why no one has asked how we calculate the 75% resp. 70% annual return of the two simple algos posted. There are different methods to calculate a return percentage, and not all are useful. Cents per share does not give any useful information about a strategy performance in my opinion.
What does "Stop = ATR(100)" mean? If it stop distance is just 1 x ATR of 100 bars, then the "trailing" stop seems to be too tight? 1 ATR is a very small number, right?
Yes, this is indeed a very good observation. 1 x ATR is too tight and when you multiply the stop distance by a higher factor, you'll get many more winning trades and the system performance jumps to almost 100%. This especially happens with the second algorithm I posted here, with the highpass filter. I just used too tight stops for giving a certain other gentleman who follows this thread another opportunity to point out the awfulness of trading with filters .
Or here is another approach: take a well worn strategy, then slightly improve one of it's aspects and hope you can extend the window of opportunity (i.e., until everyone else starts using that updated algo). Perhaps using an LPF is potentially an example like this since perhaps not many are using it and it is (at least the second order version) slightly different from the EMA. HFT is also exactly this. It is playing the spread or even a flavor of arbitrage, but with faster and faster time frames. All this is doing is giving this old strategy a little more life. But it's literally a matter of "time" before the absolute limit of low-latency is hit, and the edge will disappear.
You made some very good points. I agree with most. It is not curve-fitting in itself that does the damage but the non-stationarity that path dependent systems exhibit. Another way I will be exploiting soon is to find systems with very high win rate and low statistical significance and just fade them down to 50% win rate: http://www.elitetrader.com/vb/showthread.php?s=&threadid=215283&perpage=6&pagenumber=4
Yes, that would be like mean reversion -- it is likely to work especially for insignificant systems but you just never know when that streak will end.
I might add that a potential reason why such streaks last longer than one might expect is may be because of all the curve fitters out there?