World Cup mascot thumbnail ICML 2026 Time Series Forecasting Multimodal Counterfactual Diffusion

What if Tomorrow is the World Cup Final?Counterfactual Time Series Forecasting with Textual Conditions

Shuqi Gu1, Yongxiang Zhao1, Baoyu Jing2, Kan Ren1†
1ShanghaiTech University · 2University of Illinois at Urbana-Champaign
†Corresponding author

Paper Code Benchmark Results What-if Cases
01 · Abstract

Forecasting beyond what actually happens

One observed history can lead to many plausible futures. TADiff makes those futures controllable with natural language.

Time series forecasting is increasingly shaped by forthcoming events rather than historical patterns alone. Traditional models usually rely on historical observations or factual future conditions, making it difficult to reason about counterfactual scenarios such as unexpected traffic, weather, health, or market events.

We introduce counterfactual time series forecasting with textual conditions, where future conditions are expressed as flexible natural-language descriptions. TADiff uses a text-attribution mechanism to separate mutable textual effects from immutable historical factors, then forecasts with a condition-aware diffusion process.

The project also includes an evaluation framework for both factual and counterfactual settings. DTTC-I measures consistency with intrinsic historical factors, while DTTC-E measures consistency with future textual semantics when the true counterfactual future is unavailable.

What-if text

Flexible future conditions

Describe events in natural language instead of fixed intervention categories.

Attribution

Intrinsic history first

Extract condition-independent features before generating the future.

Diffusion

Distributional forecasts

Denoise future sequences under target textual assumptions.

DTTC

No-ground-truth scoring

Evaluate semantic consistency for futures that never happened.

02 · Method

Overview

TADiff first attributes condition-invariant historical signals, then generates factual or counterfactual futures under target textual conditions.

TADiff framework with attribution and forecasting pipelines

Training jointly optimizes attribution and forecasting. Inference first attributes xh,T from history, then uses it as the initial diffusion state for future forecasting.

1

Attribute history

Diffuse the historical sequence under historical text to remove extrinsic context and obtain intrinsic features xh,T.

2

Forecast with future text

Initialize the future diffusion state with xh,T, concatenate clear history, and denoise under the future textual condition cf.

3

Finetune for what-if shifts

Construct alternative future conditions and optimize semantic alignment even when counterfactual ground-truth futures are absent.

Diffusion

Uncertainty-aware forecasting

DDIM-style denoising models stochastic time-series evolution under given text conditions.

Attribution

Intrinsic initial state

Instead of random Gaussian noise, TADiff starts future generation from attributed historical features.

Evaluation

Factual + counterfactual

MAE/MSE are used with ground truth; DTTC handles unobserved counterfactual futures.

03 · Evaluation

Main results

16 / 20Best / Tied-best
1.50Average Rank ↓
100 75 50 25 0 93.8 TADiff (Ours) 70.0 PatchTST 55.0 Sundial 51.9 TimeMMD 48.1 IATSF 45.6 DLinear 41.9 CT 33.8 VerbalTS 28.1 TimeCMA
04 · Showcase

Showcase

05 · Citation

BibTeX

@inproceedings{gu2026tadiff,
  title     = {What if Tomorrow is the World Cup Final? Counterfactual Time Series Forecasting with Textual Conditions},
  author    = {Gu, Shuqi and Zhao, Yongxiang and Jing, Baoyu and Ren, Kan},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}