Past-error correction reduced cold bias but gave no clear accuracy gain for a DMI forecast of ENSO in 1981-2024

We tested whether past forecast errors correct a cold bias in next-year ENSO forecasts that use autumn Indian Ocean Dipole data. We used fixed HadISST1.1 DMI and CPC ERSSTv5 Nino 3.4 indices. Fits and corrections used only complete earlier targets. In 44 winters from 1981 to 2024, a ten-error correction reduced the mean cold error from 0.424 to 0.067 degrees C. Mean squared error fell by 3.42%, with a conditional 95% three-year block interval of -18.56% to 19.61%. The corrected DMI model had 14.07% lower error than the corrected Pacific model, with an interval of 3.03% to 24.15%. However, correction made the Pacific model worse. Gains against raw Pacific and a training mean were uncertain. The 25-winter subset also gave an uncertain added DMI gain. Thus lower mean error did not establish improved accuracy. The package extends a prior campaign benchmark with delayed error correction, all forecasts, source hashes, code and uncertainty draws. Short records, revised observations, simple models and conditional uncertainty limit the result. No causal ocean effect is identified.

ENSO · Indian Ocean Dipole · forecast calibration · bias correction · chronological validation