Christiansen & Nygaard evaluate SPiCT for Greenland halibut in NAFO Division 1A Uummannaq; final SPiCT model configuration accepted to provide advice

Christiansen & Nygaard evaluate SPiCT for Greenland halibut in NAFO Division 1A Uummannaq; final SPiCT model configuration accepted to provide advice

Evaluation of surplus production modelling in continuous time (SPiCT) for Greenland halibut in NAFO Division 1A Uummannaq

Northwest Atlantic Fisheries Organization

www.nafo.int Serial No. N7744

NAFO SCR Doc. 26/024

SCIENTIFIC COUNCIL MEETING –JUNE 2026 Evaluation of surplus production modelling in continuous time (SPiCT) for Greenland halibut in NAFO Division 1A Uummannaq Henrik Christiansen1, Rasmus Nygaard1 1Greenland Institute of Natural Resources, Box 570, 3900 Nuuk, Greenland 2026-06-04 CHRISTIANSEN, H., & NYGAARD, R., 2026. Commercial data from the Greenland halibut fishery in Uummannaq fjord. NAFO Scientific Council Research Document, SCR Doc. 26/024: 1-16. Abstract The Greenland halibut stock in NAFO Division 1A Uummannaq has previously been qualitatively assessed using index-based information. First attempts to provide analytical assessments with surplus production models were made since 2024. Here, sensitivity analyses are presented, i.e., an evaluation of several runs with surplus production models in continuous time (SPiCT) where input priors for r and B/K were varied while other model setup parameters were held constant. In addition, variations of truncated input catch time series were evaluated. These model evaluations were discussed in plenary and indicate a level of robustness in parameter estimates that provide an improved quantitative perception of the state of the stock compared to the previous qualitative assessment. A final model configuration is presented, which was accepted during the 2026 Scientific Council June meeting to provide advice. Introduction The inshore Greenland halibut stocks in NAFO Division 1A have been separated from the offshore Greenland halibut stock in Subarea 0+1 in 1994. The latter is assessed using a stochastic surplus production model in continuous time (SPiCT) as developed by Berg & Pedersen (2016) since 2024 (Nogueira et al., 2024). Advice for the inshore stocks in Division 1A has previously been provided without analytical assessment model, using index-based information only. Caution must be taken when applying surplus production models for these inshore Greenland halibut stocks as they are believed to rely on recruitment from the offshore stock. However, application of a SPiCT model may at least provide quantitative estimates of the exploitable biomass status and thus enable the application of the NAFO precautionary approach (PA) framework. An important consideration is then how sensitive the model reacts to the specification of different priors in terms of growth rate (r) and the initial state of the biomass of the stock in relation to its carrying capacity (B/K). Furthermore, if input time series vary in length, it should be explored whether this affects model results. Here, several (>30) model runs are evaluated and compared to investigate model sensitivity and develop an appropriate model configuration for the assessment of the Greenland halibut stock in Division 1A inshore Uummannaq.

2 Northwest Atlantic Fisheries Organization

www.nafo.int Material and Methods Input data Catch data has been available since 1954, but the first years were characterized by very low or zero catches (Nygaard & Christiansen, 2026a). Since 1977 catches have gradually increased, except for one year, 1986, with missing data. This time series from 1977-2025 was used as input for the SPiCT model. Fisheries-independent surveys were conducted in the Uummannaq area since the 1960s but only results from 1993 are currently presented due to uncertainty about earlier survey results (Nygaard & Christiansen, 2026b). Furthermore, the survey design changed from longline to gillnet in 2005 and several years in the time series are missing. An additional 90 mm mesh size section was added to the gillnet survey in 2015. Therefore, the time series from 2015 to 2025 was used as input for the SPiCT model. Details about the survey design are available in Nygaard & Christiansen (2026b). In addition, two commercial CPUEs are available: one calculated from logbook data (“commercial logbook CPUE”) and one calculated from factory landings data (“commercial factory CPUE”). The first is based on haul- by-haul information delivered to the Greenland Fisheries and Hunting Control Authority (GFJK) by fishing vessels. The latter is based on daily individual landings data that are recorded at the fish processing factories in the area. Both data sources contain metadata such as date, gear, position, effort, and more. The commercial logbook CPUE is available from 2006 to 2025, while the commercial factory CPUE is available from 2012 to 2025. Details about the CPUE calculations can be found in Nygaard & Christiansen (2026a). All index input time series (survey CPUE, commercial logbook CPUE, and commercial factory CPUE) were shifted to the middle of the year for the model runs. Base model configuration All SPiCT runs were conducted using R package “spict” v1.3.9. All model runs used the four input time series described above. The production curve was fixed to the Schaefer model, a prior for the standard deviation of the biomass was set as logsdb = c(log(0.1), 0.5), and the alpha and beta priors were disabled. No prior for the initial state of the biomass (B/K) was applied. Initially, the prior for r was set at logr = c(log(0.2), 0.2). Otherwise default settings were applied. The exact R code to run the base model is available in Appendix A. Sensitivity analyses A series of model runs was conducted, which differed from the base model run in two parameters: the r prior was varied between 0.05 and 0.35 and a prior for the initial state of the biomass as a function of the stock’s carrying capacity (B/K) was introduced and varied from 0.2 to 1 (Table 1). In total, this represents 35 individual runs with every possible combination of these tested values for r and B/K. Model runs were then checked for convergence and adherence to the points listed in Mildenberger et al. (2019). Matrices comparing the different parameter combinations were created with information about Akaike Information Criterion (AIC), ΔAIC, the biomass estimated at the end of the current year (B2025), estimated K, estimated r, estimated maximum sustainable yield (MSY), and estimated relative biomass (B/Bmsy). Catch time series truncation The catch time series used is 29 years longer than the longest used index time series. Therefore, SPiCT models using a shorter catch time series were tested. The base model configuration for this test was the same as described above, except that r was set as logr = c(log(0.25), 0.25). Two additional runs with a truncated time series were conducted. The first used a catch time series from 1991 to 2025, the same time period as used in the offshore Greenland halibut stock assessment (Hedges & Christiansen, 2025). The second used a catch time series from 2006 to 2025, thus with the same starting year as the longest used CPUE index time series. As in the sensitivity analyses above, the parameters B2025, estimated K, estimated r, estimated MSY, and estimated B/Bmsy were compared, as well as adherence to the checklist points of Mildenberger et al. (2019).

3 Northwest Atlantic Fisheries Organization

www.nafo.int Final model configuration After evaluation of the above-mentioned analyses a final model configuration was found. The long catch time series (1977-2025) was used and no prior for the initial state of the biomass (B/K) was imposed, rendering the final model setup nearly identical to the base model setup with the notable exception that the r prior was now set as logr = c(log(0.25), 0.25). The R code for the final model setup is available in Appendix B. Results and discussion All model runs converged successfully. The different runs are compared in more detail below. Sensitivity analyses Model runs were checked for adherence to the points listed in Mildenberger et al. (2019), although some of these criteria are qualitative evaluations and thus cannot be reported in a simple categorical table. However, all models passed the simpler checks, i.e. models converged, all variance parameters were finite, the main variance parameter intervals did not span more than one order of magnitude, and initial values did not influence the parameter estimates. The one-step ahead residuals varied to some extent between runs, as did the retrospective patterns. Production curves remained relatively similar between runs. The model with r prior = 0.2 and B/K prior = 0.2 had the lowest AIC value, closely followed by the model with r prior = 0.25 and B/K prior = 0.2 (Table 2). Eight model runs were within +/- 2 AIC of the model run with the lowest overall AIC (Table 3). In these eight model runs B2025 ranged from 27,815 t to 43,094 t (Table 4) and B/Bmsy ranged from 0.57 to 0.71 (Table 5). Importantly, maximum sustainable yield (MSY) varied relatively little, from 4,705 to 6,338 t in these eight model runs. But even at higher B/K priors MSY did not exceed 6,337 t (Table 6). These results therefore indicate a certain stability in parameter estimates that are relevant for providing advice in the NAFO PA framework. Catch time series truncation The catch time series was truncated to evaluate how this affects model estimates. For these three runs (catch time series start at 1977, 1991, and 2006, respectively) the r prior was relaxed again and set at 0.25. All three model runs converged, showed finite variance parameters, main parameter intervals did not span more than one order of magnitude, and initial values did not influence the parameter estimates. MSY was very similar between these three runs, ranging from 6,701 t (1977 start) to 6,965 t (1991 start, Table 7), indicating again a degree of stability in a crucial metric. Residuals and retrospective analyses showed acceptable results in all three runs. However, Mohn’s Rho was smaller in the catch time series from 1977 compared to the 1991 time series for both B/Bmsy and F/Fmsy (Table 8). In addition, the time series from 1991 estimated unrealistically high r at 0.31 (Table 7). It was therefore concluded that the time series starting in 1977 is the most promising model configuration. Final model configuration The final model configuration was thus using the catch time series from 1977, a relaxed (±0.25) r prior at 0.25 and no initial B/K prior. The final model configuration was thus using the catch time series from 1977, a relaxed (±0.25) r prior at 0.25 and no initial B/K prior. This final model estimated median relative B/Bmsy at 0.6 in 2025 (Fig. 1, Table 9) and median relative F/Fmsy at 1.9 in 2025 (Fig. 2, Table 9). Estimated parameters and relative reference points from the model are listed in Table 9. The prior and posterior distributions of the estimated parameters n (fixed to 2, i.e. Schaefer production curve), r (intrinsic growth), and sdb (standard deviation of the biomass) is given in Fig. 3. The posterior r is shifted slightly compared to the prior r. The production curve was generally characterized by the large fluctuations that the stock has experienced in the past decade (Fig. 4). The final SPiCT model showed slight residual patterns (p = 0.026 and p = 0.049 for the catch ACF and Shapiro quantiles, respectively) in the commercial CPUE from factory-landings data (Fig. 5), but slight deviations do not necessarily invalidate model results (Mildenberger et al., 2019). A five-year retrospective analysis was

4 Northwest Atlantic Fisheries Organization

www.nafo.int performed also on the final model configuration. As in all performed model runs, the model is sensitive to peels of ≥3 years, but retros are well within confidence intervals. Mohn’s rho is at 0.058 for B/Bmsy and at -0.118 for F/Fmsy (Fig. 6). The signal observed when peeling off three or more years is likely