Probabilistic models improved by combining separate predictions into joint forecasts
Interweaving Marginals into Multivariate Sample Paths: Training-Free Dependence Construction for Probabilistic Time Series Foundation Models
Machine LearningArtificial Intelligence
Summary
When predicting multiple things over time, models often give separate forecasts for each variable, without explaining how these variables work together. The authors found a way to combine these separate predictions into joint scenarios without retraining the model. This approach helps better capture how different variables relate across time and improves the usefulness of the forecasts. They show this can be done as a separate step after the original predictions are made.
What this means in practice
- •For financial analysts: Generate multivariate risk scenarios by combining independent forecasts into joint paths without retraining models.
- •For energy grid operators: Create joint future power demand and supply scenarios from separate forecasts to improve grid management decisions.
Authors
Jinmyeong Choi, Jinkwan Jang, Seul Lee, Taesup Kim
Abstract
Probabilistic time series foundation models (TSFMs) provide coordinate-wise predictive distributions, but these marginals do not determine a joint distribution over multivariate future trajectories. We study training-free coupling of frozen TSFM marginals into multivariate forecast sample paths. Our primary evaluation fixes the empirical marginal sample multiset at every channel--horizon coordinate across methods, isolating the effect of coupling alone. Historical temporal and channel relations substantially improve their corresponding dependence diagnostics. The same pattern persists when the fixed-marginal constraint is removed and paths are sampled directly, and remains present under native multivariate backbone inference. These results support treating dependence reconstruction as a distinct post-processing problem for probabilistic TSFMs.