Skip to main content
Back

VLBM Beats Every Forecasting Baseline by 15% on Rare Real-World Events

VLBM uses variational latent decomposition to keep forecasts reliable during rare, high-impact distribution shifts — without sacrificing normal-case accuracy.

  • VLBM is a new variational latent forecasting model that explicitly separates stable in-distribution dynamics from rare OOD deviations using a learned low-rank latent basis.
  • It achieves 15.08% MAE and 7.74% MSE gains over the strongest baseline across 12 benchmark tasks spanning traffic, weather, and power systems.
  • A key technical contribution is aligning a future-aware training posterior with a future-blind test prior, eliminating the train/test latency gap common in variational models.
  • The authors release a new OOD benchmark dataset aligning PeMS Bay traffic data with real California Highway Patrol incident records.
  • Unlike most robustness methods, VLBM does not trade normal-case accuracy for OOD robustness — it improves both simultaneously.
  • Code is open-source at github.com/leijieruilq/VLBM_OOD_forecast.

Most time series forecasting models are trained to be good on average. That works fine when the world behaves normally, but the moments that matter most in production , a traffic accident, a sudden power surge, an extreme weather event , are precisely the moments those models fail. A new paper from Xudong Zhang et al. introduces VLBM (Variational Latent Basis Model), a forecasting framework built from the ground up to handle both normal and out-of-distribution (OOD) conditions without trading one off against the other.

The problem with average-case training

Out-of-distribution (OOD) events are data points that look nothing like what a model saw during training , think a highway sensor during a multi-car pileup versus a normal Tuesday commute. These events are rare, but they dominate real-world risk. The trouble is that standard training optimizes for average loss across all examples. Since OOD events are rare, their gradient signals get swamped by the much larger volume of normal in-distribution (ID) data, so a model can score impressively on benchmarks while completely falling apart when it matters.

This is not a niche edge case. Real-world traffic, weather, and power systems all contain rare but high-impact OOD scenarios, and existing benchmarks do not adequately test for them. The field has largely optimized for ID accuracy, leaving a gap in OOD robustness that VLBM directly targets.

The core idea: split stable from unstable

VLBM's central insight is structural: instead of asking a single model to handle both normal and anomalous dynamics, explicitly separate them in a shared latent space. The model learns a latent basis , a low-rank subspace (think of it as a compact dictionary of the most common temporal patterns) that captures stable, recurring ID dynamics. Every input is then decomposed into two parts: