A Statistical Extended-Range Tropical Forecast Model Based on the Slow Evolution of the Madden–Julian Oscillation

Abstract

In this study, a statistical model is developed that exploits the slow evolution of the Madden–Julian oscillation (MJO) to predict tropical rainfall variability at long lead times (i.e., 5–20 days). The model is based on a field-to-field decomposition that uses previous and present pentads of outgoing longwave radiation (OLR; predictors) to predict future pentads of OLR (predictands). The model was developed using 30–70-day bandpassed OLR data from 1979 to 1989 and validated on data from 1990 to 1996. For the validation period, the model exhibits temporal correlations to observed bandpassed data of about 0.5–0.9 over a significant region of the Eastern Hemisphere at lead times from 5 to 20 days, after which the correlation drops rapidly with increasing lead time. Correlations against observed total anomalies are on the order of 0.3–0.5 over a smaller region of the Eastern Hemisphere. Comparing the skill values from the above OLR-based model, along with those from an identical statistical model using reanalysis-derived 200-mb zonal wind anomalies, to the skill values of 200-mb zonal wind predictions from the National Centers for Environmental Prediction’s Dynamic Extended Range Forecasts shows that the statistical models appear to perform considerably better. These results indicate that considerable advantage might be afforded from the further exploration and eventual implementation of MJO-based statistical models to augment current operational long-range forecasts in the Tropics. The comparisons also indicate that there is considerably more work to be done in achieving the likely forecast potential that dynamic models might offer if they could suitably simulate MJO variability.