Abstract <p>Based on satellite altimetry data on the velocity of geostrophic currents for each day from 2000 to 2023, the trajectories of passive tracers regularly distributed over a grid in the Bering Sea are calculated. The Lagrangian indicators—the lengths of the propagation paths of these tracers (L) and the Lyapunov exponent (Λ)—accumulated for the month in the past prior to reaching the sites of scientific bottom trawling (BT) have been found. It is shown that the “random forest” machine learning method, among other bagging and boosting methods, relates the logarithm of the cod density in scientific BT (t/km²) with the Lagrangian indicators in the best way; it describes more than 51% of the variance in the validation set. The vector autoregressive spatiotemporal model, standard for calculating the dynamics of cod biomass in the West Bering Sea zone, describes 6% less variance in identical testing, and the generalized additive model describes 20% less variance. All tested models in their optimal configurations included a significantly positive effect of L and a nonlinear effect of Λ in addition to the known dome-shaped effect of the depth of the BT site and the threshold effect of water temperature at the bottom.</p>