Tim Scarfe, Alexey Chernov, Wouter M. Koolen & Yuri Kalnishkan Incomplete (working) PaperThis paper introduces a new regression algorithm for on-line prediction. The algorithm is based on the following idea. A number of relatively small regions in the past are selected as training sets and instances of regression algorithms are trained on them. The resulting algorithms are then merged using techniques of prediction with expert advice. A new combination of sleeping experts and tracking the best expert is employed. The method thus constructed avoids both dealing with large matrices and retraining on the sliding window on every step. We apply the method to predicting implied volatility of options. |

Papers >