A big problem when working with models for financial prediction is the estimation of out-of-sample performance for the obtained models or trading rules. In particular, it is easy to jump into conclusions regarding trading rules that exhibit extremely profitable behavior, when tested on historical data. These misjudgments are often caused by the rules covering too few examples in the examined data. This paper deals with the problem in conjunction with nonconvex global optimization of trading rules by adding a constraint in the problem formulation. The effect is a regularization, where solutions covering too few examples are rejected. The modeling is performed with a sliding-window technique and generates different parameters for the optimized trading rules in each time window. The results from the Swedish stock market show superior generalization ability in terms of risk-adjusted hit rates for the rules generated with the proposed method. Furthermore, the results show that the high hit rates achieved, to a large extent are a result of the adaptive modeling with sliding windows. |