Increasingly powerful algorithms, such as chatbots with human-seeming capabilities, are changing the world. But can they make long-range weather forecasts?
Standard forecasts based on computer models of the atmosphere can accurately predict only a few days ahead. This is because any slight measuring error leads to growing inaccuracies, known as the butterfly effect from the way the flapping of an insect’s wings could theoretically generate a tornado. Long-range predictions are usually made in very general terms.
“It is not possible to predict the weather on a particular day months to years ahead,” states the Met Office on its page on long-range predictions.
The forecast service Weather 20/20 claims to have a better approach, using machine learning to identify recurring weather patterns and predict events months in advance. This is a new spin on the old technique of analogue forecasting in which current conditions are matched with previous weather cycles.
Weather 20/20 says it has predicted specific weather events, for example forecasting Hurricane Ian months before it formed. It is not the first to make this type of claim, and anyone can get lucky with a one-off prediction, but its apparent success has attracted considerable attention and criticism. Close scrutiny of further predictions will show if an algorithmic revolution in long-range forecasting really is on the way.
