An Error Analysis of Daily MOS Temperature Forecasts: Laying the Groundwork for Bias Correction

The spatial-temporal features of daily Model Output Statistics (MOS) weather forecast errors from 6- to 120-hour projections are analyzed. A striking feature of these daily weather forecast errors seen in the MOS guidance are relatively large-scale structures that are dominated by low-frequency variations, such as annual cycle, trend and ENSO resemble mode etc., namely climate biases. By traditional definition, climate is the average of weather. This study shows that climate biases can impact daily weather forecast accuracy, a perfect example which demonstrates a weather-climate linkage.

Further analysis demonstrates that most of these forecast errors or climate biases originated from the National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS). That is, the MOS forecast errors are inherited from errors in the predictors from the GFS model output. Although MOS can greatly reduce the GFS forecast errors, these climate biases remain either because the training samples used to build MOS forecast equations are not long enough or these biases cannot be removed by traditional MOS techniques. However, it is very encouraging that a large portion of these climate biases can be removed by using simple bias correction methods. It also appears that these bias correction techniques can be used to correct an existing MOS system for changes to the underlying version of the GFS, as new training samples are collected to update the existing statistical system.