ForecastGAN: A Decomposition-Based Adversarial Framework for Multi-Horizon Time Series Forecasting

TL;DR

ForecastGAN tackles the gap between short-term and long-term time series forecasting by integrating decomposition, horizon-aware model selection, and conditional adversarial training. The three-module architecture processes mixed numerical and categorical features, leveraging a Look-Back Window Aggregator and linear model variants tuned via validation, then stabilizes predictions with GAN-based training to yield probabilistic forecasts. Across eleven real-world datasets, ForecastGAN delivers substantial short-term gains while remaining competitive on longer horizons and offering significant computational advantages over transformer-based methods. The work provides a practical, scalable framework for diverse forecasting tasks with clear guidance on hyperparameters and deployment considerations.

Abstract

Time series forecasting is essential across domains from finance to supply chain management. This paper introduces ForecastGAN, a novel decomposition based adversarial framework addressing limitations in existing approaches for multi-horizon predictions. Although transformer models excel in long-term forecasting, they often underperform in short-term scenarios and typically ignore categorical features. ForecastGAN operates through three integrated modules: a Decomposition Module that extracts seasonality and trend components; a Model Selection Module that identifies optimal neural network configurations based on forecasting horizon; and an Adversarial Training Module that enhances prediction robustness through Conditional Generative Adversarial Network training. Unlike conventional approaches, ForecastGAN effectively integrates both numerical and categorical features. We validate our framework on eleven benchmark multivariate time series datasets that span various forecasting horizons. The results show that ForecastGAN consistently outperforms state-of-the-art transformer models for short-term forecasting while remaining competitive for long-term horizons. This research establishes a more generalizable approach to time series forecasting that adapts to specific contexts while maintaining strong performance across diverse data characteristics without extensive hyperparameter tuning.

Figures (8)

• Figure 1: ForecastGAN architecture diagram (Decomposition module has the time series decomposition element, model selection module performs model selection on four of the available models, and adversarial training module is a cGAN model with a deterministically selected Generator explained in section \ref{['sec:Decomp_module']}, section \ref{['sec:Model_module']} and \ref{['sec:Adversarial_module']} respectively)
• Figure 2: Structure of cGAN
• Figure 3: Embedding method for Decomp-Agent: Each of the continuous features is decomposed in trend and seasonality components and the categorical features are encoded (one-hot) whereas the dotted block represents the values of these features at the same time step embedded into a vector
• Figure 4: Performance comparison of different models for Long-term time series forecasting
• Figure 5: Performance comparison of different models for multi-horizon time series forecasting