In 256 RAM Legacy fish stocks, productivity regime models did not forecast 10 later years of surplus production better than a 10-year mean, and drift without steps gave the same regime frequency

Question: Vert-pre et al. (2013, PNAS) found that step changes in surplus production (regimes found with STARS) were the best models for 69% of fish stocks. If these regimes are real and persistent, the last regime must forecast later production better than models without regimes. Approach: We used RAM Legacy v4.66 (TBbest, TCbest). We removed equilibrium years, deterministic series and production-model assessments. For 256 stocks, we fitted the four Vert-pre models (Fox, STARS regimes, mixed, random; AICc) to all years except the last 10, and we forecast those 10 years from the observed biomass. We compared the forecasts with the long-term mean and a 10-year mean. We also ran the same pipeline on simulated series without discrete steps. Results: In training years, AICc selected a regime model for 69% of stocks, as in the paper. The last-regime forecast had a lower error than the long-term mean (geometric mean MSE ratio 0.71, 95% CI 0.59-0.84), but the median ratio was 1.00. A 10-year mean gave 0.67 (0.57-0.78) and was better for 61% of stocks. The last regime was not better than the 10-year mean (1.06, 0.95-1.19). For stocks with the regimes class, it was worse (1.19, 1.01-1.40). Its advantage over the long-term mean stopped after lead 3. Simulated production without steps gave regime classes for 15% of stocks with independent errors, 47% with AR(1) errors, and 57-67% with gradual drift. Under drift, the last regime was slightly better than the 10-year mean (0.95). In the data it was worse than in 98-100% of drift replicates. Results held in 9 of 10 sensitivity runs. With only 5 test years, the last regime was better than the 10-year mean (0.88, 0.77-0.99). Conclusion: production levels change, and recent levels forecast better than long-term means. But the step form did not add value at 10-15-year horizons, and autocorrelation or drift explains most regime detections. Limitations: all values are assessment outputs; test years have the largest revisions; forecasts use observed biomass; we wrote STARS in Python.

fisheries · surplus production · productivity regime shifts · STARS · RAM Legacy Stock Assessment Database · out-of-sample forecast · nonstationarity · Fox model · model selection · AICc · red noise · fish stocks