I'm not sure you read my first post in reply to you regarding statistical testing. If you need 25 billion backtests, with 500+ trades in-sample, i'm not sure you can do that on a single CPU in less than a month. Or can you?
I decided to change things up a little and instead of finding patterns that perform well when exiting 5 bars into the future, I searched for patterns that were good at holding the position until the end of the trading session. Used the same inputs as the last pattern. One thing I learned from this exercise is that these inputs are not very well suited for that type of strategy (or that type of strategy is inherently hard to trade). With 780 trades, the pattern still had a positive expectancy of +$40 per contract on ES. For more details, or to download the pattern go to <a href="http://thestrategictrader.com/technical-analysis/trading-strategy/rsi-pattern-strategy-eod-exits/812">RSI Pattern Strategy</a> I believe I have exhausted these original inputs, its time to try some new inputs and or exit combinations. Does anyone have any suggestions on inputs they would like to see? I saw a post before about using EMA inputs. I can try those next. Suggestions on how to use the inputs? For example, feed raw EMA values? EMA-Close? Or to better normalize for using the patterns against multiple instruments, we can do (EMA-Close)/EMA Thoughts?
I agree, if I added a lot of inputs and tried to determine the significant inputs requiring 25 billion backtests as you described, would require significant CPU power. For this exercise, I am choosing my inputs in a more random (suggestion based) approach instead of exhaustively testing hundreds at a time. I do like the more exhaustive approach, but I do not believe its 100% necessary to extract meaningful patterns. I believe the current approach can accomplish that goal with significantly less CPU resources. I already have one pattern that I believe is very robust. Nearly every instrument I trade it against does very well. It has also performed great in the out of sample and past 3 months of walk forward testing. Now I just need to find 100's more just like it, to use as part of a bigger strategy.
I suggest for example : John Ehlers Sine Wave Crossover I used to believe that too. Now i don't, so i think you will need an exhaustive approach, so thats just my advice. Maybe you reach the same conclusion, maybe you don't, who knows eh?
Is frost seriously trying to develop strategies using RSI? I was mistaken, thought the debate was more advanced.
RSI is just a derivative of price. When searching for a pattern you can not feed in pure price information. It needs to be adjusted/normalized in order to be useful. RSI is a good indicator in that sense. So not sure what "more advanced" your looking for. This method has already proven to find one solid pattern that can be used as part of a trading strategy.
braincell, Maybe you are successful with this approach of testing billions of strategy ideas and then going out of sample and selecting the best 50 etc etc. Personally, I do not like the idea of throwing 1 billion bombs on Afghanistan, carpet bombing it hoping one will take out Laden. I rather believe using marines or cruise missiles (=human brain) and specific intelligence inputs(=specific market behavior) to get to Laden. As we all know Laden was finally found in Pak-land and not in Afghanistan. Hope you get the analogy. So, as you can see I am firmly with ssrrkk about the approach to successful systematic trading. You mentioned many hedge funds are using methods to screen billions of strategy. I disagree with it. All the hft firms in the world work on very specific market microstructure issues and build strategies around them. They don't just randomly try some combination of RSIs and ADXs etc. However, I like to keep my eyes and ears open and always welcome criticism. So let me ask you couple of questions: 1. Do you use this specific billions of strategies datamining approach in your strategy building (or do you work for a fund that does it or do you know anyone specific fund that is doing it ?). Or is the answer no, and this is only your hypothesis that extensive datamining is a good way to build profitable strategies ? 2. If the answer is yes, I would be very interested in knowing your Sharpe and PF that you have succeeded in achieving using this extensive datamining technique. The point is - builidng trading strategies around specific market behavior, you can get a PF of 2+ and a sharpe of 3+. So, I would like to know if your approach of extensive datamining actually can give you a better Sharpe and PF. Thanks for enlightening me !
1. I don't backtest billions of random strategies to create a good strategy. I backtest billions of them to find good input values. Then i backtest/evolve via machine learning only a million with those results in mind (choosing best input values), and it turns out to be profitable. 2. There are various results, and yes such PF and sharpe happens a few times in out of sample data. I am still not satisfied though as i consider even OOS data to be in-sample to the extent that i may have manually selected those that perform best. I'll make a claim of success after a few years of live trading, until then, i just think i have a statistically significant and/or an interesting result. Also, i'll PM you a bit more info.
If such PF and Sharpe happens every now and then and at least breaks even otherwise your livetrading should be real successful?
This general topic does not seem to be garnering much interest. I think most people may believe this type of pattern searching is not valuable. In an effort to try and show that this method can generate good results going forward, I will be utilizing one of these patterns to trade S&P 500 stocks on daily bars. For tomorrow, I will be looking to enter Symbol:GAS around market close. For more information on this strategy,check out my blog at <a href ="http://mystockanalysis.com">Daily Stock Analysis</a>