We revisit the generalized inverse and ordinary demand systems for fish using four decades
of monthly Japanese seafood data (1985–2025), a period over which the market shifted from domestic
catch to heavy import dependence. The earlier consensus no longer holds. The ordinary
specification is rejected under every instrument set, but the inverse specification is also rejected
when the full set of instruments is used, so the preferred model depends on which instruments
are chosen. Additionally, lag and macroeconomic instruments that once identified the supply
side have also lost most of their explanatory power. Together these results indicate that the
structure of the market has changed and that a single demand model may no longer describe the
market well. Faced with the model selection uncertainty, we examine whether combining the
competing forecasts is preferable to selecting one. Using unconditional rolling-window forecasts
and a range of combination methods, we find that no single model forecasts best at every
horizon, while the combinations are at least as accurate as the individual models and never as
poor as the weakest among them. The gains from combination are small, because the demand
systems are very similar yielding similarl forecasts. Combination therefore offers a reliable
hedge against choosing the wrong model, even where it does not clearly improve on the best one
