CAST: Causal Anchored Simplex Transport for Distribution-Valued Time Series

2026-05-05 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

CAST (Causal Anchored Simplex Transport) is a novel causal forecasting model designed for distribution-valued time series, which are common in systems like queue occupancies, mobility shares, and public-health mixtures. These systems are observed as aggregate distributions on a probability simplex, not scalar trajectories. CAST addresses the "latent transition-kernel aliasing" problem, where similar observed distributions can arise from different underlying mechanisms, leading to inaccurate forecasts by fixed-summary methods. The model operates by retrieving empirical successors from causal context, stabilizing them with a persistence anchor, and applying a bounded local stochastic transport on ordered supports, ensuring simplex preservation at each stage. Evaluated on eleven public and simulated benchmarks spanning diverse domains, CAST achieved an average rank of 1.27 on one-step KL and 1.91 on autoregressive rollout JSD, outperforming 15 baselines by winning 8/11 sections on each metric and ranking top-2 on all 11 sections for offline KL.

Key takeaway

For AI Scientists and Machine Learning Engineers working with distribution-valued time series, CAST offers a robust solution to the challenge of latent-kernel aliasing and ordered mass motion. You should consider implementing CAST, especially for applications like queueing systems or public health, where preserving the simplex and accurately modeling local mass shifts are critical for stable, long-horizon predictions. Its strong performance across diverse benchmarks suggests it can significantly improve forecast accuracy and reliability.

Key insights

CAST improves distribution-valued time series forecasting by structuring transitions around causal context, persistence, and local mass transport.

Principles

• Transition operators should be structured around the simplex.
• Latent-kernel aliasing incurs irreducible excess risk for fixed-summary forecasters.
• Ordered supports benefit from bounded local stochastic transport.

Method

CAST retrieves empirical successors, forms a persistence-retrieval anchor $a_{t}=\lambda_{t}p_{t}+(1-\lambda_{t})r_{t}$, and applies a bounded local stochastic transport $T_{t}$ to yield $\hat{p}_{t+1}=(1-\rho_{t})a_{t}+\rho_{t}T_{t}a_{t}$.

In practice

• Use CAST for forecasting queue occupancies and air-quality severity profiles.
• Apply CAST to ecological compositions and energy generation mixes.
• Consider CAST for risk-sensitive planning where full distribution forecasts are critical.

Topics

• Causal Anchored Simplex Transport
• Distribution-Valued Time Series
• Latent-Kernel Aliasing
• Probability Simplex
• Stochastic Transport

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• A Stationary (and Therefore Compatible) Representation is All You Need — Stationary representations learned via d-Simplex fixed classifiers inherently ensure model compatibility over time.
• PLOT: Progressive Localization via Optimal Transport in Neural Causal Abstraction — PLOT localizes neural causal variables efficiently using optimal transport, reducing computational overhead for mechanistic interpretability.
• LazyDINO: Fast, Scalable, and Efficiently Amortized Bayesian Inversion via Structure-Exploiting and Surrogate-Driven Measure Transport — LazyDINO offers fast, scalable, and efficiently amortized Bayesian inversion using derivative-informed neural surrogates and lazy maps.
• Distribution Alignment for One-Shot Federated Learning via Optimal Transport — SLOT-Align harmonizes heterogeneous client feature distributions in One-Shot Federated Learning using optimal transport for improved accuracy.
• Triangular-Reference Schr\"odinger Bridges for Time Series Generation — TR-SBTS enhances time series generation by using a hierarchical, adaptive volatility reference for improved fidelity.
• ARIA: Adaptive Region-Based Importance Allocation for Conditional Diffusion Distillation — ARIA adaptively allocates training effort in conditional diffusion distillation by tracking teacher-student discrepancy across conditioning regions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.