In our quest for ways to predict markets or stocks, the idea of Walk Forward Testing arises from time to time. Basically, the assumption is that if you examine your data, say 1, 2, 3 years of data, you may find a strategy that can predict future price movements. To do this, you divide the data into segments. For example, if you have 3 years of data, you might divide it into the oldest 2 years, and the most recent 1 year. Theory has it that if you find something that worked on the oldest two years, it should work on the most recent year. In this example, the older two years of data is called “out-of–sample” data, and the most recent data would be called “in-sample” data.
You can do the same fragmenting of data using months, or even weeks. The idea is the same. Divide the data into ‘in-sample’ and “out-of-sample” parts, find something that works on the “out-of-sample” portion, and hope that it works on the “in-sample” part.
Having done a lot of this sort of thing, I suggest that, bottom line, it doesn’t work. I believe there is a better way. In my experience, your results will be closer to reality, if you back test using 3 months, 6 months, 9 months and one year. If you find a strategy that works in these time frames, it’s likely to work for the coming 3-6 months going forward. Any further testing is unnecessary, in my opinion.