Every year, farmers in southern India wait for the same thing: the sky to open up and the monsoon to arrive. Whether rice sprouts and whether there is enough water for the whole season depends on it. So the monsoon forecast is not abstract science but a matter of survival for hundreds of millions of people. The CFSv2 model, used by meteorologists, has long been considered especially accurate when started in February. But new research shows that this accuracy is an illusion, born of a single lucky strike.

To understand what went wrong, picture a marksman aiming at a target, but his hand trembles. On one shot, the tremor accidentally compensates for the aiming error, and the bullet hits the bull's-eye. If you judge only by that shot, the marksman seems a master. Similarly, in 1983 the model predicted La Niña, a cooling of the tropical Pacific, which did not actually happen. But this erroneous prediction coincidentally matched the real excess rainfall caused by a completely different mechanism: the positive phase of the Equatorial Indian Ocean Oscillation. One "lucky" miss boosted the average score of February forecasts.

When the researchers removed 1983 from the calculations, February's advantage melted away. Moreover, other metrics, such as mean error, bias, and root mean square error, had pointed from the start to April and May giving more reliable forecasts. The problem is not statistics but the model's physics. It ties rainfall too strongly to Pacific temperatures: in the model, the correlation between the NINO3.4 index and rainfall reaches -0.85, whereas in real observations it is only -0.44. The model seems to think the monsoon is driven solely by the Pacific, ignoring the Indian Ocean's own variability. As a result, it incorrectly links EQUINOO to ENSO and even flips the sign of the relationship between EQUINOO and rainfall: in the model the correlation is -0.77, while in observations it is +0.54.

This physical bias means that in years when the monsoon is determined by the Indian Ocean rather than the Pacific, the model is systematically wrong. And if you look only at correlation, these errors can be masked by random coincidences, as in 1983. Operational centers should reconsider their choice of initial conditions: April–May starts give more stable results, and above all, the model itself needs fixing, reducing its excessive sensitivity to El Niño and the cold bias in the central Pacific.

The marksman in our analogy does not become a master from one lucky shot. Likewise, the model does not become better from one random coincidence. To trust a forecast, you need to look not at individual hits but at how the marksman performs over a long distance, and then the February monsoon forecast turns out to be not a gold standard but a trap for the unwary.