This paper evaluates the accuracy of linear, non-linear time series models as also both models combined in forecasting major macroeconomic variables - Monthly series of Index of Industrial Production (IIP) and quarterly series of Gross Domestic Product (GDP) - in Indian context.
The paper observes that for IIP (and its sub component) series, in the short horizon (1-6 months), forecast combination (median) is found to be marginally better performing than linear as well as non-linear modelling framework. In the long horizon (7-12 months), non-linear models perform relatively better than linear models as well as combination forecast. For GDP (and its sub-component) series, forecast combination using median forecast, has been found to be performing relatively better for both short horizon as well as long horizon.
Additionally, the paper observes that the accuracy of the forecast improves by using combination forecast for series with long memory property/less volatile series.