>How do you back test it to know that it works. If you back test over the past 5 years then you are only testing your model against a huge bull market.
>If you back test over the past 20 years then I'm not sure it helps much as the market of 20 years ago didn't really have any of the major market drives of today's markets, HFT's, huge numbers of hedge funds and the money they bring, and huge passive investing via ETF's, all those factors were there 20 years ago, but they weren't the major market drivers that they are now.
I imagine you wouldn't want to 'back test' over some arbitrarily selected number of years of data. You'd want an adaptive algorithm that could exploit 'trends' over any time period.
But, like someone else mentioned, backtesting is near meaningless: the only thing that matters is actual performance.
> You'd want an adaptive algorithm that could exploit 'trends' over any time period.
Of course that what you want, but how will you find it, and more importantly, how will you know you found it? In a bull market, an "adaptable general algorithm" that makes 10% in a year is statistically indistinguishable from an algorithm that buys and holds S&P500, unless you design your backtests very carefully.
>If you back test over the past 20 years then I'm not sure it helps much as the market of 20 years ago didn't really have any of the major market drives of today's markets, HFT's, huge numbers of hedge funds and the money they bring, and huge passive investing via ETF's, all those factors were there 20 years ago, but they weren't the major market drivers that they are now.
I imagine you wouldn't want to 'back test' over some arbitrarily selected number of years of data. You'd want an adaptive algorithm that could exploit 'trends' over any time period.
But, like someone else mentioned, backtesting is near meaningless: the only thing that matters is actual performance.