Researchers at the Stevens Institute of Technology studied the performance of seven short-term (nowcasting) models over the New York City area for 2014 to 2022. In the study, they used five deterministic models and two probabilistic models, along with radar data at various lead times to determine the accuracy of the various models.
Deterministic models, which assume that local precipitation cells (such as thunderstorms) will not change over time, were found to be best for very short-term forecasts up to 20 minutes from now. Although both the deterministic and probabilistic forecast models were highly accurate in this short range, the probabilistic models take longer to generate a forecast. Those models take longer than 20 minutes to process and run, rendering them useless for those very short-term forecasts. After that time period, probabilistic models–which take into account the changing nature of these cells –become much more accurate and useful.
This is similar to the earliest days of numerical weather prediction models, when computational power was so limited that although models to forecast the weather were being developed, the technology did not yet exist to make them useful; by the time they had finished running, the time period for which they had generated forecasts was already passed. It continues to be an issue for forecasting how a thunderstorm or tornado, for example, may change in the next 10-15 minutes or so, as no model that takes into account dynamic changes in the thunderstorm can be run instantaneously. Thus, the best meteorologists can do is assume the storm will remain as it currently is, or make educated guesses on how it will change.
The study also found that the varying models performed better in certain situations. For example, models that were designed to forecast intense rainfall performed better in the summer months, when this type of rainfall is much more likely. It is important, therefore, to know how to select the best model for the time period, location, and situation for which one is needed.