Suppose I do an analysis and find a stock with a high Kurtosis (makes good deviations) and it is skewed to the right substantially. So I know that there are more strong movements upwards , whereas the down movement tend to take place less often and aren't perhaps as strong (even if they are, the stop loss can stop that) . So, it is a good stock to trade, but how do you go about devising the best strategy for trading? And how much data do you account for...if the stock has a long history, then prices inculcate different times, so to absorb say the data 20 years ago and look at the 'whole' would make sense? I mean how do you define the cutoff period in terms of data?
Back-testing is useful but fraught with danger. No matter how large the data set, there is always the concern that you will fall victim to the phenomenon know as "curve fitting." One of the common strategies to avoid this problem is to optimize your rules using a randomly selected part of the historical data set. Then validate your rules using the part of the data set not used in the optimization. This is only a general idea and there are many details not mentioned. A text I found useful is "Design, Testing, and Optimization of Trading Systems" by Robert Pardo published in 1992.
2 years historical data should suffice. Check the Top/Bottom Line Numbers for the previous 5 years , if Earnings Growth is achieved than check its performance Relative to the broader Index or the sector in which the stock is a constituent of.
You might prefer writing a blank check to your brokerage. At least doing so you don't waste time. You can rest assured you will lose with short term trading, no matter what strategy you use. Only very long term trading (many years) pays off.
Let's say you get good results with your latest idea in-sample... if it's random then there is maybe fifty percent chance that you will get good results out of sample... maybe it would be better to test on quite a few sets of out of sample data...
IIRC, Pardo's book goes into this. If I have some time in the next few days, I'll try to summarize his insights into your question.
You mean making long-term bets (based on economic analysis) or just that trading doesn't pay off in the short-term?
I remember doing the optimization thing for a trading course. The book that we used had optimization techniques based on some physics phenomenons. But the thing I could never understand with optimization is: Shouldn't there really be one optimum way to trade looking into the past? So, why are there many techniques that would optimize the trading returns and stop loss, all yeilding different results? Is it that one strategy exists and there are various ways to search for it which may end up finding local minima and stuff (be more or less effective in finding that strategy), or are there many such strategies inherently and different methods get to different ones?
Perhaps an analogy will illustrate the point. When I travel from Ohio to Seattle I have a number of choices. I can fly commercially. I can use my pilots license to fly myself in a small private aircraft. I can drive my car. I can take the bus. I can take the train. You get the idea. All will get me there but each has its tradeoffs with respect to speed, cost, safety, comfort, etc. There are many methods to discover rules and strategies to obtain optimum results based on your criteria. Different methods may even yield the same results. Some brute force methods are so computer intensive that it challenges even today's computing technology. Other search techniques, such as the use of genetic algorithms, are orders of magnitude faster. They find good or even great solutions, but not necessarily the "best" solution.