Integration of Judgmental and Statistical Approaches to Forecasting

When it comes to forecasting, it's important to know how good your forecasting is and if there are ways to improve it. This work focuses on finding reliable and informative indicators of forecasting performance and on how to improve forecasts with the use of judgment. Chapter 2 explores limitations of various error measures and introduces a new class of metrics (AvgRel-metrics) for measuring forecasting performance using the following rules: i) relative indicators are averaged across series using the weighted geometric mean, ii) an indicator used to evaluate forecasts must correspond to the loss function used to optimize forecasts. The AvgRelMSE and AvgRelMAE metrics are proposed to measure accuracy under quadratic and linear loss, respectively, and the AvgRelAME to measure bias. Boxplots of logs of relative indicators are used to visualize distributions. Chapters 3 and 4 look at models for handling unaided judgment & judgmental adjustments. In particular, the thesis introduces advanced models based on using panel data and Bayesian analysis. Chapter 5 proposes a novel approach allowing to incorporate judgment into a joint model and update forecasts as new data becomes available.