Answers are grounded in this paper and cite their sources.
Introduction Daily grid-scale forecasts of catch occurrence are increasingly used for operational planning and to track species redistribution under climate change, yet what sets their predictability limit is rarely quantified. Methods We assembled 6 years (2020–2025) of daily Japanese flying squid ( Todarodes pacificus ) catch, vessel-activity, and ocean-reanalysis records for Korean waters: 20.35 million grid-days on a 0.05° grid, over 99% zero-catch. Results The 2025 holdout is a genuine distribution shift rather than a modeling artifact: 32 of 33 covariates stayed below the 0.25 major-shift threshold of the population stability index while catch per vessel fell 11.9-fold inside the fixed optimal-temperature habitat. On that holdout, daily occurrence forecasting is bounded by persistence, the trivial forecast that repeats the previous day's map (one-day-ahead 8-Neighbor F1 0.506) and so encodes only the inertia of yesterday's fishing. Three improvements (a Focal-Tversky loss, a ConvLSTM–Transformer hybrid, and a validation-weighted ensemble) reached 0.347 (five-seed mean), and two designs embedding persistence scored 0.486 (five-seed mean) and 0.502 (one seed); none exceeded persistence detectably, so the bound is not an artifact of insufficient tuning. The constraint is inherited from the record rather than from model capacity: national fishery statistics measure effort as the presence of all operating vessels rather than squid-directed ones, so only 1.81% of effort-positive grid-days carry a squid catch. The vessel-presence field persists almost deterministically where the catch field does not (8-Neighbor F1 0.923 versus 0.506), and a catch in the neighboring cells on the previous day raises the probability of a catch 14.4-fold while the environmental windows used in the feature design raise it not at all. Intensity, unlike occurrence, proved only weakly determined by the environment (best R² about 0.05, Tweedie and log-normal, on the densest cells), so the two components are modeled separately. Discussion Because that inertia reflects how the statistics are collected rather than how the models are built, the bound is expected to hold for fine-scale forecasts based on records of this kind. To our knowledge, this is the first quantification of such a ceiling for daily, grid-scale catch forecasting, and it identifies squid-directed effort records as the decisive next input.
No takes yet. Share an insight, caveat, or question.
Park et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: