0 Th National Convention on Statistics (ncs) Forecasting from an Additive Model in the Presence of Multicollinearity Forecasting from an Additive Model in the Presence of Multicollinearity

In a regression model with time series data, whether the regressors are cointegrated or not, the structure of X’X is usually characterized by multicollinearity. An additive model is postulated and estimated via the backfitting algorithm. The model sequentially smooth the residuals by entering one variable-at-a-time into the equation until the residuals behave randomly or all the effects of the independent variables are estimated. Real and simulated data exhibits superiority of the method over ridge estimators and principal component regression in prediction. Furthermore, even with the poorest model fit, the method is capable of providing ‘good’ estimates of the coefficients of the variables that entered into the model early during the iterative process.