Augmented Functional Time Series Representation and Forecasting with Gaussian Processes (# W21)
Nicolas Chapados and Yoshua Bengio University of Montreal Canada

We introduce a functional representation of time series that forecasts over an unspecified time horizon (given as input variable) using progressively-revealed information sets. Use Gaussian Processes to get a complete covariance matrix between forecasts.

Forecast covariance leads to a risk-aware trading criterion. Application to actively trade price spreads between commodity future contracts. We demonstrate significant out-of-sample risk-adjusted returns after transaction costs on a spread portfolio.