Rapport Klima 15, 2002

Rasmus E Benestad:

Month-to-seasonal forecasting at met.no

Maps of monthly mean sea surface temperature and sea level pressure have been reconstructed for the 1900-2000 period. Analysis of the sea level pressure and sea surface temperature together with monthly mean station observations suggests a real relationship between certain spatial patterns and the subsequent monthly-to-seasonal mean temperature at various locations in Fennoscandia in winter. There is also a possible weak lag relationship in summer, but the results give no indication of any predictive skill in spring. Three prototype month-to-seasonal forecasting models have been tested.

The test results suggest that there may be some predictability associated with monthly mean and seasonal mean temperature. At least part of the predictability of temperature is associated with the high winter-time autocorrelation.The monthly-to-seasonal mean precipitation is generally more difficult to forecast than temperature according to the estimated skill scores.

A new strategy for developing empirical month-to-seasonal forecasting models is proposed. Historical data contain some errors which affect the calibration of the empirical models. One way to deal with errors is to use of weighting functions based on data quality, but in many cases the the quality has not been established. The proposed strategy deals with errors by selecting only a few data points, potentially avoiding bad data. Long historical records of monthly mean observations are sub-sampled iteratively by selecting different random ensembles over a large number of times. The random-ensemble-model (REM) strategy produces a large number of forecasts that can be used to estimate a forecast probability distribution (p.d.f.). The REM model has been tested with climate station records as predictor, but is designed to use any kind of data (e.g. SST fields) as input.

Key words: month-to-seasonal forecasting; empirical studies; verification