Structure aware model improves reliability of multivariate forecasts
Structure-Aware Graph Abstention for Reliable Selective Forecasting
Machine Learning
Summary
Forecasting future events can be tricky, especially when predicting several things at once that depend on each other. The authors found that existing methods sometimes give confident predictions that look okay overall but ignore important connections between variables. They developed a new way to decide when the model should say "I’m not sure" by checking both individual prediction quality and how well predictions fit together using a learned graph. This approach makes forecasting more reliable, especially when those variable relationships matter a lot.
What this means in practice
- •For demand forecasting teams: Improve the reliability of sales and inventory predictions by identifying and abstaining from uncertain forecasts involving multiple interdependent products.
- •For environmental data teams: Enhance multivariate weather or climate forecasts by selectively refraining from predictions when cross-variable relationships appear unreliable.
Authors
Jianxiang Xie, Belal Alsinglawi
Abstract
Selective forecasting abstains on high-risk test windows under a retained-coverage budget. Existing gates such as TEM (Brusokas et al., 2025) score each forecast as a whole; for multivariate outputs, trajectories can look plausible while violating dependencies among variables. We treat instance-level plausibility and relational consistency as distinct reliability axes and operationalize the latter via a learned sparse graph and a Dirichlet-style structural energy E_struct, trained with error-weighted graph regularization and score-error alignment. On seven long-horizon benchmarks and four backbones, structural gating often reduces selective MSE versus TEM at matched coverage, with the largest gains where cross-variable structure appears more informative in our benchmarks; gains are not universal, indicating a complementary abstention signal. Table 1 is a Protocol A ranking diagnostic (seed 2024); three-seed deployable Protocol B on an aligned subset is in Table 3 (full validation-to-test grids: Appendix A).