Development and Evaluation of a Linear Inverse Model for Weeks 3-4 Forecasts of 2-Meter Temperature Over North America

Abstract:

This talk presents an empirical-dynamical model designed for weeks 3-4 forecasts of 2-meter temperature over North America. In support of the Climate Prediction Center’s week 3-4 outlook, we have trained a year-round linear inverse model (LIM) on global observations from the past 40 years. LIMs employ a linear approximation of the predictable dynamics of a system (including the linear parameterization of rapidly decorrelating nonlinearities) constructed from the statistics of the system itself. The model state vector is formed by the leading principal components of time-averaged variables, focusing on slowly-evolving large-scale features for which processes can be approximated as linear, and predictable. The state vector includes 7-day average anomalies of 2-meter temperature, sea level pressure, 500-hPa height, tropical heating, and stratospheric circulation. The skill of the model is evaluated for forecasts of 2-meter temperature anomalies and their categorical probabilities. Cross-validated hindcasts are compared with IFS and CFS hindcasts. Weeks 3-4 skill of the LIM and these operational models is comparable except during winter when LIM Week-3 skill is much poorer. We will present the seasonality and spatial distribution of skill scores across North America for the common hindcast periods.

Additionally, while current operational dynamical models exhibit low mean subseasonal skill, individual forecasts can be identified with greater than average skill. LIMs have been shown to be useful for identifying such “forecasts of opportunity”, by predicting its own expected skill from the signal-to-noise ratio. We show that the skill of the major operational dynamical models is also sorted by the LIM expected skill. As such, the LIM can identify forecasts of opportunity in which higher confidence can be expressed in subseasonal outlooks. These forecasts of opportunity exhibit a large increase in local anomaly correlation and Heidke skill score, in addition to larger spatial coverage of high probabilities for the categorical outlook.