This study compared the predictive performance of a traditional multivariate time series model, the Vector Autoregression model, and a machine learning-based XGBoost model for the mid to short-term forecasting of coastal fishery catches. The results showed that while the VAR model, which accounts for seasonal volatility and lag effects in historical time series data, exhibited relatively high predictive volatility, the XGBoost model demonstrated more stable predictions by learning nonlinear patterns and complex inter-variable relationships. Notably, in the XGBoost model, lagged variables of past catches and derived variables related to aquaculture production were identified as important predictors, confirming that coastal fishery catches are significantly influenced by short-term volatility and aquaculture production factors. Although both models showed similar overall trend directions, differences were observed in the magnitude of fluctuations. The VAR model generally indicated a declining trend, replicating past patterns in its forecasts, while the XGBoost model showed a more gradual and stable decline.
KIM et al. (Fri,) studied this question.