A New Look at Weather Forecasting through Analogues

In the literature, the use of analogues for short-range weather forecasting has practically been discarded. This is because no good matches for today's extratropical large-scale flow patterns can be found in a 30-year data library. We propose here a limited-area model approach for Analogue-Forecasting (AF). In order to make a 12-hour AF valid at a target point, we require analogy in initial states only over a circle with radius of about 900 km. On a limited area there are usually several good analogues, sometimes to within observational error. Different historical analogues may be used at different target points. The usefulness of the limited area approach is first demonstrated with some examples. We then present verification statistics of 3000 12-hour 500-mb height point forecasts in the Northern Hemisphere winter at 38°N, 80°W (over West Virginia, U.S.A.). In order to beat persistence at 12 hours at this point we need an analogue which differs by about 40 geopotential meters or less from the base case. This requirement is met almost all of the time using a 15-year dataset for analogue searching. We find a few percent of the analogue pairs to be within observational error. In the mean, over 3000 cases the initial discrepancy is 33 gpm. When averaging over the first five analogues 12-hour AF over the eastern United States can be characterized by a 52 gpm rms error and 0.95 (0.77) anomaly (tendency) correlation. The forecasts have the correct amplitude, i.e., no damping, in spite of the averaging over five individual forecasts. We then show an example of a 500-mb height forecast map on a (roughly) 2000 × 2000 km area over the eastern part of North America. Although different analogues were used to arrive at the 12 hour forecast at each of the 25 gridpoints, the resulting map looks meteorological and the forecast is moderately successful. A verification of a large set of 12-hour forecast maps shows that the height gradients are indeed forecast with some skill. We then proceed to make 24-hour point forecasts by finding historical limited-area matches to the 12-hour forecast maps. This second time step indicates that the AF-process holds up, with forecast accuracy increasing its margin over persistence. Two applications are discussed. Comparing initial and 12-hour forecast error tells us something about “error growth” and predictability at that spot according to a perfect model. Given that there are usually several good analogues, Monte Carlo experiments and probabilistic forecasts naturally suggest themselves. In particular we find the spread of analogue forecasts to be an eminent predictor of forecast skill. Refinements and applications and extension to longer range forecasts are discussed in the final section.