Meta Heuristic Approach for Automatic Forecasting Model Selection

Selection of appropriate forecasting models with their optimized parameters for a given business scenario is a challenging task and requires reasonable expert knowledge and experience. The problem of selecting the best forecasting model becomes computationally complex when the business needs forecasts on thousands of time series at a given time period. Many a times business users are interested in adapting the best parameter settings of proven forecasting models of the past and use them for further predictions. This approach facilitates the users to identify the forecasting model with a parameter value which minimizes the average of forecast errors across all the time series. This paper proposes a genetic algorithm based solution approach which simultaneously suggests the suitable forecasting model and its best parameter(s) value which minimizes the average mean absolute percentage error of all the time series. This approach is tested on randomly generated data sets and the results are compared with few randomly selected samples. For a fair comparison the samples are tested in SAS 9.1 and the results are compared with sample results which used GA suggested forecasting model and parameter values.