Cognitive Computing at the Edge: Lag Analysis and Neural Network Forecasting for Fishery Landings in the Colombian Caribbean

Fishery-dependent indices in data-limited fisheries are short and incomplete, which raises the question of when adaptive learned models, including cognitively inspired ones deployed at the edge, are justified. We analysed monthly landings per unit effort (LPUE) of pink shrimp (Penaeus notialis) in the Gulf of Morrosquillo, Colombian Caribbean: 65.7% of values are missing and the longest continuous segment has 28 months. Naive, trend, autoregressive (AR), echo state network (ESN), LightGBM and LSTM models were compared in rolling-origin experiments with training-only preprocessing, and their cost was measured on an NVIDIA Jetson Orin Nano. One-step-ahead errors were 5.1–5.7 kg per trip for the trend, the AR(2) and the learned models, against 7.6 for persistence; no learned model showed a robust out-of-sample improvement over a linear trend or an AR(2). An exploratory model incorporating river flow at a data-selected lag achieved the lowest RMSE (4.2 kg per trip), but its improvement over the statistical baselines was not statistically conclusive and was not consistently reproduced when lag selection was restricted to the training data at each origin. Backcasting was worse than persistence. A fit took 8 ms (AR) to 3–9 s (LSTM); the double-precision ESN execution process showed the largest peak memory footprint (1.7–4.1 GB of 8 GB). The limit is the information in the series, not the platform; under these conditions a linear trend or an AR(2) is a sufficient reference.