More statistical testing - a surprising result

As time passes with the trading, more data is collected and the robustness of various statistical tests improves. But intra-day trading generates data quickly. At the sample size I use - typically 10 or 15 minutes - a trading day contains up to 90 discrete time periods. I assume markets to be "self-similar" - i.e to look the same at small resolution as they do at larger. Therefore a week of intra-day data is, in some sense, the same as around two years of daily data. My trading research picked up from the end of February and so some of the systems have nearly four months of statistical data attached to them - the equivalent of 30 years of daily data. So quite a few statistical results are quite significant now

And so to the odd result that has been thown out. Basically I have been looking at the issue of portfolio construction for the last few weeks i.e. what precise combination of systems to use and in what weightings. So I am producing detailed results for each system on its own and then looking at the various blendings of these. The odd results is this - one of the systems is statistically far and away the best, and so it tends to swamp the portfolio construction process. Measured by annualized equivalent of weekly Sharpe Ratio, it produces the extraordinary number of +17. So far it is yet to have a losing week, yet returns an unleveraged average of +1.2% per week. Quite extraordinary!

Yet the system has faced a wide variety of market conditions - days when the market hardly moves at all, to days when it has a 600 point up and down range. So the result does not come from the unique trading conditions of the last four months. If the system could produce merely one-third of this result going forward it would be an incredible result.

Most importantly, tweaking the systems parameters, either the time period sampled, or the actual parameters used in the various calculations, makes very little difference to the result, which is excellent news.

So what I now have to think about is whether to accept this result at face value, in which case this system should be significantly overweight in my portfolio, or wait for yet more statistical data.