Machine-Learning for Dynamic Reverse Engineering of Hedge Funds

The leave-one-out cross-validation in nested sets of data models is traditionally considered in Machine Learning as the basic instrument of finding the most appropriate subset of features or regressors in pattern recognition and regression estimation. We extend the notion of a nested set of models onto the problem of time-varying regression estimation, which implies, in addition to the generic challenge of choosing the subset of regressors, also the inevitable necessity to choose the appropriate level of model volatility, ranging from the full stationarity of instant models in time to their absolute independence of each other. So, there are, at least, two axes of model nesting in the problem of nonstationary regression estimation, first, the relevant size of the set of regressors and, second, the level of model volatility in time. We use the leave-one-out measure of the model fit as quality indicator along both nesting axes. We apply the proposed technique to analysis of a hedge fund's returns and reverse-engineering its strategies.