Abstract
Fisheries management increasingly relies on biological reference points, target fishing mortality and biomass levels, to set sustainable catch limits. For single-species stocks these reference points can often be computed from a fitted assessment model, but when a stock’s mortality is driven by multispecies predation rather, the reference points have no closed form and depend on the assumed status of every species in the system. Obtaining them requires repeatedly re-projecting a multispecies model forward, at every assessment year of simulations, and ignoring this predation-driven mortality process is known to bias both stock assessment outputs and the reference points used for management advice. This computational cost is one of the bottlenecks in evaluating harvest strategies against multispecies operating models, including one built here for the Eastern Bering Sea groundfish complex (walleye pollock, Pacific cod, arrow tooth flounder) using Rceattle, the R implementation of CEATTLE (Climate Enhanced, Age-based model with Temperature-specific Trophic Linkages and Energetics).Here we show that a gradient-boosted tree model, guided by age-structured natural-mortality and trained on terminal-year assessment output from 500 closed-loop management strategy evaluation simulations, can emulate these multispecies reference points, without re-running the dynamic projection, achieving cross-validated R2 of 0.80–0.95 across six targets. Feature importance analysis reveals that a single pollock-specific natural-mortality feature dominates prediction for four of the six targets, even for cod’s and arrow tooth flounder’s own reference points, a machine-learned signature of the predation coupling built into the operating model. This demonstrates that these emulator models can substitute for costly re-projection, offering a path to rapid screening of harvest-strategy across the simulation ensembles that multispecies management strategy evaluations increasingly require.