Artificial skill in monsoon onset prediction: two recent examples

TL;DR

This paper critically examines the skill of monsoon onset prediction methods, highlighting biases in verification practices and emphasizing the reliability of dynamical models over trend extrapolation approaches.

Contribution

It reveals biases in current verification practices and compares the predictive skill of different methods, advocating for the use of dynamical models.

Findings

IMD operational forecasts have inflated skill scores due to data overlap.

Trend extrapolation methods show low correlation with observations (as low as 0.24).

Dynamical models demonstrate more reliable skill (~0.7) than trend-based methods.

Abstract

For two cases of empirical monsoon onset prediction it is argued that current verification practice leads to optimistically biased skill, caused by the intricacy of the model setup. For the case of the operational forecasts by the Indian Meteorological Department (IMD) it leads to an overlap of model definition and verification data. A more seriously flawed verification was used in a recent method based on trend extrapolations of 'tipping elements' (TE). Claims of TE of predicting onset 2 weeks earlier than other methods are unjustified. On the contrary, the correlation between TE forecasts and observations is as low as 0.24 and compares poorly to the reported IMD correlation of 0.78. That latter value likely being artificially inflated, currently the best and most reliable monsoon onset predictions come from a dynamical model with more reliable skill values of about 0.7.