State-Dependent Merit-Order Effects in a Hydro-Dominated Market: Conditioning Variable-Renewable Price Suppression on the Marginal Fuel Regime in New Zealand

The merit-order effect (MOE), the suppression of wholesale prices by zero-marginal-cost variable renewable energy (VRE), is typically estimated as a single average coefficient per market. In a hydro-dominated system, the price-setting technology alternates between hydro and thermal plants, so the MOE should depend on the operating state. Using New Zealand nodal data, a baseline panel model first reproduces a published 2011–12 seasonal benchmark to within NZ$0.22/megawatt-hour (MWh). A model conditioning the VRE coefficient on reservoir storage (2015–24, 245 nodes, 28,827 node-month observations from approximately 42 million half-hourly records) passes every in-sample diagnostic (within R2 = 0.81) yet fails out-of-sample and stability testing. The interaction adds no significant predictive value beyond a storage main effect (p = 0.15–0.22) and reverses sign across sub-periods (+58 to −51), an instability not attributable to collinearity. Replacing the storage level with the monthly thermal generation share, a dispatch-based proxy for the hydro-thermal regime, yields a stable, significant, out-of-sample-validated interaction (clustered t = 5.0, leave-one-month-out cross-validation gain +14.2%) that survives direct hydrological and demand controls, including a head-to-head test against a storage interaction. Identification rests on 120 monthly observations, since these regressors are system-wide. Price suppression is strongest in hydro-abundant months, at approximately NZ$5.8/MWh per 10% relative VRE increase, and attenuates toward zero as the thermal share rises, consistent with water-value pricing. A forward test on Jan 2025 to May 2026 data, which postdate model development, confirms the interaction’s predictive value (forward RMSE improvement +33%, wild cluster bootstrap p = 0.047). Re-estimation on the 2011–12 panel returns a weakly identified, directionally opposite interaction (bootstrap two-sided p = 0.053). The validated state dependence therefore characterises the 2015–24 market regime rather than a structurally invariant relationship.