Predictive models for financial data are often based on a large number of plausible inputs that are potentially nonlinearly combined to yield the conditional expectation of a target, such as a daily return of an asset. This paper introduces a new architecture for this tasK: On the output side, we predict dynamical variables such as first derivatives and curvatures on different time spans. These are subsequently combined in an interaction output layer to form several estimates of the variable of interest. Those estimates are then averaged to yield the final prediction.
展开▼
