Title: Enhancing Precision and Variety: A Multi-Task Strategy for Merging Forecast Models

Abstract: This study introduces a multi-task optimization framework grounded in deep learning for time series prediction. By utilizing extensive time series datasets, the proposed method determines the optimal weights for combining various forecasting models to generate predictions for individual series. This approach simultaneously tackles two core objectives: choosing distinct forecasting models and effectively merging them. Uniquely, the method balances both the precision and the diversity of the selected forecasting techniques.

The architecture comprises two distinct modules. The model combination module retrieves specific features to optimize the weights assigned to each forecasting method. Concurrently, the model selection module identifies alternative features to pinpoint the subset of models suitable for prediction. This selection process is formulated as a classification task, where the labels correspond to the specific group of models to be employed for a given series. These labels are derived from an auxiliary optimization problem designed to locate methods that offer both high accuracy and diversity for each time series.

The outputs from both modules are integrated, and the entire neural network undergoes joint training through gradient descent, minimizing a bespoke loss function. Testing on a substantial collection of series from the M4 competition dataset, as well as real-world road traffic data, demonstrates that our approach improves point forecast accuracy relative to existing state-of-the-art techniques.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC