📊 statistics

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

This paper introduces CAST, a causal forecasting method for distribution-valued time series that leverages simplex-structured successor retrieval and anchored stochastic transport to overcome latent transition-kernel aliasing, achieving state-of-the-art performance across diverse real-world and synthetic benchmarks.

Original authors: Jiecheng Lu, Jieqi Di, Runhua Wu, Yuwei Zhou

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiecheng Lu, Jieqi Di, Runhua Wu, Yuwei Zhou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Predicting the "Shape" of Things, Not Just Numbers

Imagine you are trying to predict the future of a busy coffee shop.

Old way (Scalar): You just guess, "Tomorrow, we will sell 150 cups." This is a single number.
New way (Distribution): You guess the entire mix of what will happen: "There's a 20% chance of a slow morning, a 50% chance of a steady rush, and a 30% chance of a chaotic line where people leave."

The paper argues that many real-world systems (like traffic, air quality, or computer server queues) don't just move up and down like a single number. They shift their shape. The "CAST" model is a new tool designed specifically to predict these shifting shapes.

The Problem: The "Look-Alike" Trap

The authors identified a major problem with current forecasting tools called Latent Transition-Kernel Aliasing. Let's call this the "Look-Alike Trap."

Imagine two different days at the coffee shop.

Day A: The shop is currently empty (0 customers). But, a massive bus just dropped off 50 people outside, and they are about to flood in. The future will be chaotic.
Day B: The shop is currently empty (0 customers). But, it's a holiday, and no one is coming. The future will be quiet.

If you only look at the current state (0 customers), a standard AI might say, "Well, usually when it's empty, it stays empty," or it might average the two outcomes and predict "a little bit of chaos." It gets confused because the current picture looks the same, but the history is totally different.

The paper claims that standard models fail here because they average these different futures into a blurry, useless guess.

The Solution: CAST (Causal Anchored Simplex Transport)

The authors built a new model called CAST. Think of it as a detective that doesn't just look at the crime scene (the current moment) but investigates the history to solve the case.

CAST works in three simple steps, like a three-part recipe:

The Memory Search (Retrieval):
Instead of guessing, CAST looks back at its own history. It asks: "Have I seen this exact situation before? If so, what happened next?"
- Analogy: If you see a dark cloud, you don't just guess "rain." You check your memory: "Last time I saw this cloud, it was a Tuesday and it rained hard. The time before, it was a Saturday and it was just a breeze." CAST finds the specific "look-alike" moments from the past to see what actually happened next.
The Anchor (Stabilization):
Sometimes, the past isn't a perfect match. To avoid jumping to wild conclusions, CAST creates a "safety anchor." It blends the current situation with the retrieved memory.
- Analogy: It's like saying, "Okay, the past suggests a storm, but right now the wind is calm. Let's bet on a mix: mostly calm, but with a chance of rain." This keeps the prediction grounded and prevents it from flying off the rails.
The Local Move (Transport):
This is the secret sauce for things that have an order, like "Low," "Medium," and "High" traffic. CAST knows that mass (people, cars, data) usually moves to neighboring spots, not teleporting across the map.
- Analogy: If a crowd is moving from "Low Traffic" to "Medium Traffic," CAST predicts they will step into the "Medium" bin. It doesn't predict they will suddenly teleport to "High Traffic" or "Zero Traffic." It respects the natural flow of the crowd.

Why It Matters: The "Simplex"

The paper uses a fancy word, Simplex, but you can think of it as a Pie Chart or a Percentage Bucket.

In many systems, the total must always equal 100%. (e.g., The percentage of energy coming from Solar + Wind + Coal must equal 100%).
Standard AI models often mess this up, predicting 110% or negative percentages.
CAST is built with a "mathematical guardrail" that ensures the prediction is always a valid pie chart. It never breaks the rules of percentages.

The Results: How Did It Do?

The authors tested CAST on 11 different real-world datasets, ranging from:

Ecology: How different animal species share a habitat.
Energy: How much electricity comes from wind vs. solar.
Health: The mix of causes of death in a population.
Traffic & Queues: How many people are waiting in line.

The Verdict:
CAST won the most competitions. It was the best at predicting the future shape of these systems in 8 out of 11 categories.

Short-term: It predicted the next step better than almost everyone else.
Long-term: When they let the model predict 64 steps into the future (a "rollout"), CAST stayed stable and accurate, while other models got blurry or drifted off course.

The "Aliasing" Experiment

To prove their theory, the authors created a fake, controlled experiment where two different futures looked exactly the same at the start.

Standard models failed miserably, averaging the two futures and getting the prediction wrong every time.
CAST looked at the hidden history, figured out which "regime" (scenario) was active, and predicted the correct future with near-perfect accuracy.

Summary

CAST is a new forecasting tool that understands that history matters. It doesn't just look at the current snapshot; it remembers similar past moments, blends them with the present, and moves the prediction in a logical, local way. This allows it to predict the shifting "shapes" of complex systems (like traffic or weather) much better than older methods that just guess a single number or average out the possibilities.

Technical Summary: CAST (Causal Anchored Simplex Transport)

1. Problem Statement

The paper addresses the forecasting of distribution-valued time series, where observations are not scalar trajectories but probability distributions evolving over time. These distributions live on the probability simplex $\Delta^{D-1}$ and represent aggregate system states such as queue occupancies, mobility shares, air-quality severity profiles, and ecological compositions.

The core challenge identified is latent transition-kernel aliasing. In many stochastic systems, the same observed current distribution $p_t$ can arise from different latent regimes (e.g., a queue entering congestion vs. draining after a burst) with distinct future successors. Traditional forecasters that rely solely on a fixed summary of the current state (or current-only inputs) are forced to average these incompatible successors, leading to an irreducible error. Furthermore, standard deep learning architectures (Transformers, RNNs) treat these inputs as unconstrained vectors and enforce simplex validity only via a final softmax, failing to explicitly model the causal structure of successor locality, regime disambiguation, or ordered mass motion.

2. Methodology: The CAST Model

The authors propose CAST (Causal Anchored Simplex Transport), a structured transition operator designed specifically for the simplex. CAST operates as a causal, successor-local pipeline with three distinct stages, all of which preserve the simplex by construction:

Causal Successor Retrieval:
CAST encodes the causal history $p_{1:t}$ into hidden states using a decoder-only Transformer. It then retrieves empirical successors $r_t$ from a strictly causal memory bank of past context-successor pairs $(p_s, p_{s+1})$ where $s < t$ . This retrieval mechanism allows the model to disambiguate regimes: if the current state $p_t$ has appeared in different contexts with different futures, the retrieval mechanism selects the successor corresponding to the current causal context.
Persistence–Successor Anchor:
To stabilize the noisy retrieved successor, CAST forms an anchor $a_t$ by interpolating between the current distribution $p_t$ and the retrieved successor $r_t$ :
$a_t = \lambda_t p_t + (1 - \lambda_t) r_t$
Here, $\lambda_t$ is a learnable persistence gate. This step ensures the forecast remains close to the current state if the retrieved history is unreliable, while still leveraging historical transitions when appropriate.
Bounded Local Transport (for Ordered Supports):
For supports with a natural order (e.g., queue lengths, AQI levels), CAST applies a bounded local stochastic transport operator $T_t$ to the anchor. The final prediction is:
$\hat{p}_{t+1} = (1 - \rho_t) a_t + \rho_t T_t a_t$
where $\rho_t$ is a bounded transport strength. The transport kernel $T_t$ is constrained to move probability mass locally (e.g., radius-1 shifts) and includes a "mean-shift budget" to prevent runaway drift during autoregressive rollouts. For unordered supports, this stage is disabled ( $\rho_t \equiv 0$ ).

Training: The model is trained using teacher-forced one-step KL divergence loss. Crucially, it does not use rollout loss, multi-horizon targets, or oracle supervision. A target-free operator prior ( $R_{op}$ ) is added to penalize strong transport, off-identity mass, and rough kernels, ensuring stability without direct supervision on future steps.

3. Theoretical Contributions

The paper provides three main theoretical results regarding the limitations of fixed-summary forecasters and the expressivity of CAST:

Aliasing Lower Bound (Theorem 1): The authors prove that under latent transition-kernel aliasing (where multiple regimes share the same observed summary but have different successors), any fixed-summary forecaster incurs an irreducible weighted Jensen–Shannon (JS) excess-risk lower bound. This bound is strictly positive unless all regime-specific successors are identical.
Oracle Representability (Theorem 2): They demonstrate that the CAST hypothesis class contains the regime-aware Bayes successor. If the causal representation can identify the latent regime and the memory contains the appropriate anchor, CAST can theoretically achieve zero excess risk.
Ordered-Support Separation (Theorem 3): For ordered supports, the authors show that even an oracle method restricted to the "no-transport" anchor class (linear combinations of current and retrieved states) incurs a conditional KL excess risk bounded below by a Pinsker separation term if the true successor lies outside the anchor hull. CAST, by utilizing bounded local transport, can bridge this gap.

4. Experimental Results

The authors evaluated CAST on a suite of 11 benchmarks spanning ecology, energy, diet, mortality, employment, air quality, severe weather, mobility, and queueing (both G/G/1 and Gt/G/1).

Performance: CAST achieved the best average rank on both one-step KL (1.27) and autoregressive rollout JSD (1.91) across all 11 sections.
Comparisons: It outperformed a broad baseline set including statistical methods (persistence, ETS, ilr-VAR), compositional models, and modern deep time-series architectures (Transformers, iTransformer, N-HiTS, TimeMixer, etc.). CAST won 8 out of 11 sections on both metrics and was top-2 on all 11 for offline KL.
Ablations: Removing the persistence mix caused a significant drop in performance (KL wins dropped from 8/11 to 5/11), while removing structural regularization primarily degraded long-horizon rollout stability.
Synthetic Validation: A controlled synthetic experiment confirmed the aliasing theory. Fixed-summary forecasters collapsed to the JS-average of incompatible successors, while CAST successfully recovered the regime-specific successor, closing the error gap to near zero.
Qualitative Analysis: Visualizations on queueing data showed that while baselines produced static bands or smeared distributions, CAST preserved the moving modes and shape of the distribution, consistent with its bounded transport mechanism.

5. Significance and Claims

The paper claims that CAST offers a structural alternative to generic sequence-to-distribution models. Its significance lies in:

Explicit Modeling of Causal Structure: By retrieving empirical successors and anchoring them, CAST directly addresses the aliasing problem where current states are insufficient to determine futures.
Simplex-Native Design: Unlike models that map to unconstrained logits and project back, CAST's transition operator is constructed to remain within the simplex at every stage, ensuring valid probability distributions without post-hoc correction.
Stability in Rollout: The combination of persistence anchoring and bounded local transport prevents the accumulation of errors during long-horizon autoregressive forecasting, a common failure mode for deep time-series models.
Regime Disambiguation: The results suggest that for distribution-valued time series, leveraging causal history to distinguish latent regimes is more critical than simply increasing model capacity or sequence length.

The authors remain modest regarding the scope, noting that the theory provides an existence argument for the regime where CAST excels, rather than a finite-sample training guarantee. They acknowledge that for purely smooth dynamics on unordered supports, other models may remain competitive, and that the current implementation relies on one-step training with rollout as a pure evaluation protocol.