Deterministic reservoir computing for subseasonal rainfall forecasting under model-structure misspecification: a computer-science evaluation of forecast-value loss in rainfed advisory services

Subseasonal rainfall forecasts underpin rainfed agricultural advisories, yet the forecast systems that reach advisory services are typically evaluated on statistical accuracy rather than on the economic value of the decisions they inform. The central problem is model-structure misspecification: when the reservoir's assumed dynamical form departs from the generating process, forecast errors are not merely larger but differently structured, and the mapping from error structure to decision loss is nonlinear. The paper formalises forecast-value loss for a threshold-triggered advisory, derives a bound relating reservoir spectral radius and readout conditioning to the divergence between forecast-error and decision-loss orderings, and shows that two reservoirs with identical mean squared error can produce materially different advisory value. A methodological protocol is proposed for evaluating deterministic reservoir forecasters against decision-relevant criteria rather than accuracy alone. The contribution is a formal and methodological reframing of reservoir-computing evaluation for climate advisory services, with implications for how forecast value is reported to agricultural extension systems.