🤖 machine learning

EventTSF: Event-Aware Non-Stationary Time Series Forecasting

EventTSF is an autoregressive diffusion framework that improves non-stationary time series forecasting by integrating textual event data through event-aware flow-matching timesteps, thereby addressing the challenges of multimodal interaction and imbalanced denoising difficulty to significantly outperform existing baselines.

Original authors: Yunfeng Ge, Ming Jin, Yiji Zhao, Hongyan Li, Bo Du, Chang Xu, Shirui Pan

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yunfeng Ge, Ming Jin, Yiji Zhao, Hongyan Li, Bo Du, Chang Xu, Shirui Pan

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

Imagine you are trying to predict the weather for next week. A traditional weather forecaster looks only at the temperature, wind, and rain from the past few days. They might say, "It looks like it's going to rain because it rained yesterday." But they miss the big picture: a massive music festival is happening downtown tomorrow, which will cause traffic jams, crowds, and a sudden spike in local humidity that the past data alone couldn't predict.

This is the problem the paper EVENTTSF tries to solve.

The Core Problem: The "Blind" Forecaster

Most time series forecasting models (used for energy, traffic, stock markets, etc.) are like that blind weather forecaster. They look at the numbers but ignore the story behind them.

The Issue: Real-world data is "non-stationary," meaning the rules change constantly. A sudden event (like a holiday, a news headline, or a power outage) can completely change how the data behaves.
The Gap: Existing models either ignore these text-based events or try to mix them in clumsily, resulting in predictions that are smooth and safe but often wrong when things get chaotic.

The Solution: EVENTTSF (The "Story-Aware" Predictor)

The authors built a new system called EVENTTSF. Think of it as a super-smart detective who doesn't just look at the crime scene (the numbers) but also reads the police report (the text events) to understand what happened.

Here is how it works, using simple analogies:

1. Synchronizing the Story with the Numbers

Imagine you have a long video of a busy street (the time series) and a script describing what's happening in the video (the text events).

Old Way: The model watches the video and reads the script separately, then tries to guess the future. They never really "talk" to each other.
EVENTTSF Way: It cuts the video into short clips and locks each clip to the specific sentence in the script that describes it. If the script says "Parade starts," that sentence is glued to the exact moment the crowd surges in the video. This allows the model to learn exactly how a specific type of event changes the numbers.

2. The "Denoising" Challenge (The Dirty Window)

The model uses a technique called Diffusion, which is like trying to clean a window covered in thick fog to see the picture underneath.

The Problem: In the past, models cleaned the window using a fixed schedule. They assumed every part of the picture was equally hard to clean.
The Reality: Some parts of the picture (like a calm day) are easy to clean. Other parts (like a chaotic event day) are covered in thick, sticky mud and are very hard to clean.
The Innovation: EVENTTSF introduces an "Event-Aware Timestep."
- Imagine a smart cleaning robot. If it sees a calm day, it cleans quickly. If it sees a chaotic event (like a riot or a sale), it slows down and spends extra time scrubbing that specific spot.
- The model looks at the text event first, realizes "Oh, this is a chaotic day," and adjusts its cleaning schedule accordingly. This prevents the model from getting confused or giving up on the hard parts.

What Did They Find?

The authors tested this new detective on 7 different datasets, ranging from synthetic (made-up) data to real-world traffic, weather, and atmospheric physics.

The Results: EVENTTSF crushed the competition. It outperformed 12 other top models.
- It improved probabilistic forecasting (guessing the range of possible outcomes) by 41.3%.
- It improved deterministic forecasting (guessing the single most likely outcome) by 27.5%.
The "What If" Test: They tried adding text events to the old models to see if that helped. It helped a little, but not nearly as much as EVENTTSF. This proves that just having the text isn't enough; you need a model built specifically to understand the relationship between the story and the numbers.

A Note on "Bad" Stories

Interestingly, the paper found that the model didn't work well on one specific dataset: Traffic News.

Why? The news articles were too long and full of fluff (background noise) that didn't actually relate to the traffic. It was like reading a 50-page novel to find out if it's going to rain in 10 minutes. The model got confused because the "story" didn't match the "action."
Lesson: The text events need to be clear, relevant, and concise for the model to work its magic.

Summary

EVENTTSF is a new way to predict the future that treats text events (like holidays, news, or promotions) as essential clues, not just background noise. By syncing the story with the numbers and adjusting its learning speed based on how chaotic the event is, it makes much more accurate predictions than models that only look at the numbers.

Technical Summary: EVENTTSF – Event-Aware Non-Stationary Time Series Forecasting

1. Problem Definition

Time series forecasting is critical for sectors like energy and transportation, yet it faces persistent challenges due to non-stationarity and distribution shifts. While recent advances (e.g., normalization, model decomposition) have addressed single-modality non-stationarity, they often overlook the fact that distribution shifts are frequently entangled with external events from other modalities, particularly natural language texts (e.g., holidays, promotions, weather reports).

Existing approaches suffer from two primary limitations:

Modality Gap: Most methods rely on single-modal data or use coarse-grained textual context, failing to model fine-grained interactions between discrete external events and continuous time series.
Diffusion Inefficiency: Classical diffusion-based forecasting models typically sample diffusion timesteps from a fixed uniform distribution. This ignores event-induced non-stationary variability, where different events cause distinct shifts in signal magnitude and variance. Consequently, uniform sampling leads to imbalanced denoising difficulty across diffusion stages, as the Signal-to-Noise Ratio (SNR) varies heterogeneously depending on the event context.

2. Methodology: EVENTTSF

The authors propose EVENTTSF, an autoregressive diffusion framework designed to integrate historical time series and textual events via step-wise diffusion. The architecture addresses the synchronization of discrete events and continuous series, as well as the imbalance in denoising difficulty.

2.1 Multimodal Autoregressive Diffusion

Event Synchronization: The model partitions continuous time series into non-overlapping segments ( $x_s$ ) and aligns each with a corresponding textual event embedding ( $c_s$ ). If no event occurs, a NULL event is assigned. This creates a unified representation $D = \{(x_s, c_s)\}$ , bridging the gap between temporal point processes and continuous series.
Autoregressive Generation: The model treats forecasting as an autoregressive generation of fine-grained segments. At each step $s$ , it conditions on a latent state $Z_{s-1}$ (summarizing history), the current noisy segment $x^t_s$ , and the aligned event embedding $c_s$ .
Flow Matching: Instead of standard diffusion, the model utilizes Flow Matching to learn a velocity field $v_t$ that transforms a simple prior distribution into the target data distribution. The denoiser $M_\theta$ is trained to predict this velocity field, updating the latent state $Z_s$ autoregressively.

2.2 Event-Aware Timestep Sampling

To mitigate the imbalanced denoising difficulty caused by uniform timestep sampling, EVENTTSF introduces an event-aware flow-matching timestep.

Problem: Uniform sampling ( $t \sim U(0,1)$ ) fails to account for the fact that different events induce different SNRs, leading to heterogeneous denoising difficulties.
Solution: The model reparameterizes a base uniform timestep $\hat{t}$ into an event-adaptive timestep $t$ using a learnable mapping function $\phi$ :
$t = \phi(\hat{t}, c) = F^{-1}_{\text{Beta}(\alpha(c), \beta(c))}(\hat{t})$
Mechanism: Two lightweight linear networks predict the parameters $\alpha(c)$ and $\beta(c)$ of a Beta distribution conditioned on the event embedding $c$ . This ensures the effective denoising difficulty is allocated more evenly across diffusion stages for specific event types, adapting the sampling schedule to the event semantics.

2.3 Architecture Details

The core denoiser is a Multimodal U-shaped Diffusion Transformer (M-U-DiT). It employs down-sampling and up-sampling with skip connections to align event semantics with multi-resolution temporal patterns. It integrates historical states, noisy inputs, and event conditions via cross-attention mechanisms.

3. Key Contributions

Paradigm Reformulation: The paper formulates event-aware non-stationary forecasting as a multimodal paradigm, explicitly incorporating external event information to address distribution shifts that single-modal models miss.
Methodological Innovation:
- Fine-Grained Alignment: Introduces a synchronization mechanism to align discrete event segments with continuous time series windows.
- Event-Aware Timestep: Proposes a novel reparameterization strategy that learns event-conditioned diffusion schedules to balance denoising difficulty, moving beyond static uniform sampling.
Superior Performance: Extensive experiments demonstrate that EVENTTSF outperforms 12 baselines (including non-stationary and diffusion-based models) across 7 datasets.

4. Experimental Results

The model was evaluated on 7 datasets (1 synthetic, 6 real-world covering traffic, weather, and atmospheric physics) against 12 baselines (e.g., CSDI, TimeDiff, TMDM, NsDiff, Autoformer, DLinear).

Probabilistic Forecasting: EVENTTSF achieved an average improvement of 41.3% in Continuous Ranked Probability Score (CRPS) and Weighted Quantile Loss (WQL) across all datasets.
Deterministic Forecasting: The model improved Mean Absolute Error (MAE) by an average of 27.5%.
Ablation Studies:
- Removing textual events degraded performance significantly (e.g., WQL increased by 10.71% on the Temperature–New York dataset).
- Replacing uniform timestep sampling with the event-aware mechanism consistently reduced errors (e.g., CRPS reduced by 9.21%).
Efficiency: EVENTTSF demonstrated faster training times (1.13× faster than the second-best baseline on synthetic data) and sublinear growth in training time relative to forecasting horizons, attributed to its segment-wise generation approach.
Baseline Augmentation: Even when baselines were augmented with event embeddings (via cross-attention or residual conditioning), they remained substantially inferior to EVENTTSF, indicating that simple injection of events is insufficient without the specific autoregressive diffusion and timestep mechanisms.

5. Significance and Claims

The paper claims to pioneer the integration of external multimodal knowledge specifically for non-stationary time series forecasting. Its significance lies in:

Addressing the Modality Gap: It provides a unified framework for modeling the interaction between discrete textual events and continuous time series, capturing cascading and lagged dependencies that global context models miss.
Solving Diffusion Imbalance: It identifies and solves the issue of event-induced heterogeneity in denoising difficulty, proposing a learnable, event-conditioned sampling schedule that stabilizes training and improves generation quality.
Practical Applicability: The framework is designed for scenarios where future events (e.g., planned promotions, holidays) are known in advance, allowing the model to anticipate distribution shifts that cannot be inferred from historical observations alone.

The authors conclude that EVENTTSF establishes a flexible paradigm for incorporating heterogeneous event semantics into temporal modeling, opening new directions for analyzing cross-modal non-stationary dynamics.