DecoVAE Offers Lightweight, Interpretable Probabilistic Time Series Forecasting

Probabilistic time series forecasting remains a significant challenge, largely due to the difficulty in accurately modeling distinct trend and seasonal dynamics. Many existing methods struggle to capture the unique internal properties of these components, often lack interpretability, or incur heavy memory and runtime overhead. To address these limitations, researchers propose DecoVAE, a lightweight and interpretable Variational Autoencoder (VAE) framework. DecoVAE explicitly decomposes time series into its trend and seasonal components by applying domain-specific inductive biases. The trend stream enforces structural smoothness using a differential regularizer on the latent trajectory, akin to the Hodrick-Prescott filter. Concurrently, the seasonal stream operates in the frequency domain via a complex Gaussian VAE, naturally capturing the amplitude and phase of periodic patterns. Extensive evaluations across seven real-world benchmarks demonstrate DecoVAE's consistent outperformance of strong baselines. It achieves reductions of up to 14.96% in CRPS and 23.30% in NMAE for short-term forecasting, and even more substantial gains of up to 52.68% and 26.51% for long-term horizons. Crucially, these accuracy gains come with high efficiency, reducing model weight by up to 93% and accelerating speed by up to 74% compared to the second-best method.