In my automated strategy work, I am trying to better understand the relationship of outliers.
Looking at stocks for example, say 5 years of data, 10k trades spread over 20 symbols. In a particular strategy I am testing at the moment, I am seeing that less than 10% of those trades (outliers) account for 90% of the profit.
In this particular example, the strategy has no parameters. It's a moving average and an oscillator. No preset targets or stops. Market orders, commissions included. No targets on the entry bars.
I have tried tweaking the input data, for example changing the behavior from exit on close to true and false, and while the results (profit) do change, the outlier percentage does not largely change. I have tried scanning 500 stocks that are not in this basket, and I have found similar patterns in terms of outliers.
I am trying to better understand the significance of these outliers for this particular strategy. I created a frequency distribution of all 10k trades and found that after eliminating non-profitable trades, roughly 90% of the profitable trades were off low profit value, while the remaining 10% were of high value.
The purpose of the post is just to talk to someone else who has done similar work and find out if they've had similar findings. These findings do not echo my own personal discretionary trading, so I take pause. The first consideration was naturally being over fit, but I have eliminated this possibility by testing it on almost 1,000 stocks over a 5 year period and the strategy has zero parameters. There has been no optimization.
I will be forward testing this naturally, but I am trying to speed up the out of sample testing by just having a better grasp on the results. I will be on vacation for 3 weeks and will use this period to analyze out of sample results and compare them, but it is insufficient time to properly measure outliers.