Time Series Forecasting With a Learning Algorithm : An Approximate Dynamic Programming Approach

We consider the basic problem of refi tting a time series over a finite period of time and formulate it as a stochastic dynamic program. By changing the underlying Markov decision process we are able to obtain a model that at optimality considers historical data as well as forecasts of future outcomes. We design lookahead dynamic methods for the solution of our Markov decision process. By recursively applying this idea over a range of future time periods, look-ahead dynamic programming methods effectively react to changes in the data and consider the stream of future outcomes obtained from our model decisions. These techniques give rise to models calibrated to historical data which at any point in time would be optimally positioned to react to a possible future data stream