Formalizing the data-mining task
We want to find a trading strategy that will perform well over the next year.
As a proxy, we look for the strategy that performed best over a historical period.
We conjecture that this strategy will perform well in the near future. The conjecture will be true if:
- the near future is similar to the training period, and
- patterns used in the strategy are not artifacts.
But there is no reason to expect the single best historical strategy to be the best in the future.