AROpt: A Novel Optimization Method for Autoregressive Time Series Forecasting

Researchers have introduced AROpt, a new training method designed to enhance autoregressive time series forecasting. Current models, primarily based on transformer-style neural networks, often rely on scaling up model size for long-term predictions rather than utilizing genuine autoregressive rollout. AROpt addresses this by enforcing two key properties: ensuring prediction errors increase monotonically with the forecasting horizon, and enabling the concatenation of short-term predictions for flexible long-term forecasts. Violations of the error-growth trend are penalized during training to ensure consistency. Empirical results indicate that AROpt establishes a new state-of-the-art performance across multiple benchmarks, achieving a mean squared error reduction of over 10% compared to strong baselines like iTransformer. Additionally, the method allows short-horizon models to perform reliable long-term predictions at horizons more than 7.5 times longer than previously possible. The associated code has been made publicly available to facilitate further research and application in machine learning and artificial intelligence domains.