Western volume raised six-month ENSO forecast error by 10.81% against one month of history in a constant model at 2000-2024 origins

We tested whether current western Pacific warm water volume can replace one previous monthly state in a small ENSO model. We used frozen NOAA GODAS and ERSSTv5 records and 300 monthly forecast origins in 2000-2024. Each expanding least squares fit predicted six months ahead. In the constant model, a current three-state model had 10.81% higher temperature mean squared error than a two-state model with one added lag. The conditional 95% three-year block interval for the error increase was 2.91% to 19.92%. Adding the lag still cut error by 9.69% with western volume present. With monthly intercepts and current-state slopes, the primary error difference was uncertain. Five-year blocks gave the same distinction. This extends prior two-state history tests with an explicit measure of ocean structure. It does not establish physical memory or operational forecast skill. Revised inputs, model choice and a short dependent test record limit inference. Code, exact source hashes, aggregate scores, all score draws, numerical checks and a figure are included.

ENSO · western Pacific · warm water volume · memory · vector autoregression · forecast comparison