Hi, i began forward testing my strategy that i paper traded a while. My question is what do you feel is a good sample size to analyze the results before reviewing/modifying something ? Here are some details, - trading stocks - average holding time of 2.75 days - around 10 trade a week (varies) - non-automated I am shooting for 100 trades, I am playing small right now and i can afford the testing length regards, Totantaz
Sample size is not the only issue here. You also need to consider price history length. If for example most of the trades were generated during a medium term strong uptrend and they are all long you have to determine whether your results could have been obtained by a random system. For dailly bars I use about 12 years of history and my total output of signals must be in the order of 250 or more. Then I forward test on about 2 years worth of data and I expect a sample of no less than 40 trades. Then, as suggested many times in this forum one must check whether the results are random.
ah yes, thanks for the reply, i usually try to have the same number of short position than long ones. When volatility is high, the market usually drag half of my positions toward my stops but then the other half are doing well. the edge in that system are the stocks that go to my targets whatever the market is doing. For my system to work, i have to get more of those correctly than bad calls to start with. you mention backtesting years of data. I only paper traded my system for a couple of weeks before forward testing with real capital. I really want to see the effect of slippage and comms. cheers, Totantaz
you might want to read Roughtrader's postings on this subject in this tread. They could be helpfull to you
For cross-validation techniques it is standard to use 80% of data to generate strategy and 20% of data to test. However yes I agree you need to consider will your strategies still work in bull vs bear vs non-trending markets.
The ultimate issue with any trading system is whether its signals are the result of some robust timing algorithm or are essentially random. The answer(s) to the above question can be approached only with backtesting, unless one is willing to pay for it with real testing. Actually, this is the purpose of backtesting. In this other thread there is a discussion of how one can approach this problem with some examples http://www.elitetrader.com/vb/showthread.php?s=&threadid=198383&perpage=6&pagenumber=1