This website is useful for exploring backtest overfitting.
Trading in larger sizes than would be optimal. If the expected size of mis-pricing or arbitrage is 10 basis point, then no model should have been trade in such a way as to have 15 basis point of market impact. As this can not be sustainably profitable.
Adaboost:
-
First weak learner minimize the number of errors:
-
Second weak learner focus on the previously mis-classified points by punishing more on the previously mis-classified points in the cost function.
-
Third, forth, ... learners do the same thing.
