An unobserved component modeling approach to evaluate multi-horizon forecasts (Discussion Paper 2018-04)

We propose an unobserved modeling framework to evaluate a set of forecasts that target the same variable but are updated along the forecast horizon. The approach decomposes forecast errors into three distinct horizon-specific processes, namely, bias, rational error and implicit error, and attributes forecast revisions to corrections for these forecast errors. By evaluating multi-horizon daily maximum temperature forecasts for Melbourne, Australia, we demonstrate how this modeling framework can be used to analyze the dynamics of the forecast revision structure across horizons. Understanding forecast revisions is critical for weather forecast users to determine the optimal timing for their planning decisions.