Volatility Forecasting with Sparse Bayesian Kernel Models

Motivated by previous findings that discretization of financial time series can effectively filter the data and reduce the noise, this experimental study, performed in a realistic setting of trading straddles via predicting volatility, compares trading performances of symbol-based models with those of probabilistic models operating on real-valued sequences. We show that carefully designed probabilistic models trained in a Bayesian framework of automatic relevance determination can achieve superior trading performances.