Testing the Efficiency of Bias Predictions in Forecasting

Combining forecasts from multiple models into one forecast typically leads to better forecasting performance than individual models can achieve. However, elaborate backward looking combination strategies are often outperformed by simple strategies, such as averaging. Gibbs and Vasnev (2018) show that a forward looking approach based on predictable bias can make combinations that outperform equal weighting, individual and random walk forecasts. In this thesis, those results are generally replicated. The extension of their analysis is within the modification of bias predictions to have autoregressive terms to account for possible serial correlation. Forecast combinations have been made with these new bias predictions for both inflation and unemployment rate data and the results show that forecasting inflation generally benefits from adding autoregressive terms to the bias predictions, whereas unemployment forecasting performance does not improve.