Forecasting with an adaptive control algorithm

We construct a parsimonious model of the U.S. macro economy using a state space representation and recursive estimation. At the core of the estimation procedure is a prediction/correction algorithm based on a recursive least squares estimation with exponential forgetting. The algorithm is a Kalman filter-type update method which minimizes the sum of discounted squared errors. This method reduces the contribution of past errors in the estimate of the current period coefficients and thereby adapts to potential time variation of parameters. The root mean square errors of out-of-sample forecast of the model show improvement over OLS forecasts. One period ahead in-sample forecasts showed better tracking than OLS in-sample forecasts.