Technical Summary: Counterfactual Time Series Forecasting with Textual Conditions

Problem Definition

Time series forecasting in real-world scenarios is often hindered by the reliance on historical patterns and factual future conditions, which fails to account for complex, stochastic, and hypothetical ("what if") scenarios. Existing methods typically depend on historical observations or simple, structured future interventions (e.g., binary event indicators), limiting their ability to generalize to unstructured, real-world complexities. Furthermore, evaluating forecasting performance under counterfactual conditions—where the future event described in text did not actually occur and thus lacks ground-truth time series data—presents a fundamental challenge for traditional accuracy-based metrics like Mean Squared Error (MSE).

The paper formalizes the task of Counterfactual Time Series Forecasting with Textual Conditions. The goal is to predict future time series sequences conditioned on:

1. Historical time series data ($x_h$) and its associated historical textual conditions ($c_h$).
2. Future textual conditions ($c_f$) that may be factual (observed) or counterfactual (hypothetical/stochastic).

Key challenges identified include:

• Pattern Conflicts: Historical dynamics and future conditions may exhibit conflicting patterns (e.g., a World Cup Final causing traffic spikes despite typical low-traffic hours).
• Adaptability: Models trained solely on factual data struggle to adapt to unseen counterfactual combinations, often producing homogenized forecasts.
• Evaluation: The absence of ground-truth future sequences in counterfactual settings renders standard numerical metrics inapplicable.

Methodology: TADIFF

The authors propose TADIFF (Text-Attributive Time series Diffusion), a multimodal diffusion model designed to address these challenges through a novel text-attribution mechanism.

1. Text-Attribution Mechanism

To balance the influence of historical patterns and future conditions, TADIFF decouples the time series into intrinsic features (fundamental properties independent of extrinsic conditions, such as demographic structures in traffic) and extrinsic conditions (contextual factors like weather or events).

• Attribution Stage (Inference): The model uses a condition-aware diffusion forward process to "eliminate" the influence of historical extrinsic conditions ($c_h$) from the historical sequence ($x_{h,0}$), extracting the intrinsic feature ($x_{h,T}$). This serves as the initial noise state for the forecasting process.
• Forecasting Stage: The model conditions the denoising process on the original historical sequence ($x_{h,0}$) and the target future condition ($c_f$). It uses the extracted intrinsic feature ($x_{h,T}$) as the starting point for generating the future sequence, ensuring the forecast respects the underlying intrinsic dynamics while adapting to the new textual condition.

2. Training Strategy

• Joint Optimization: The model is trained with two objectives:
  • Attribution Loss ($L_A$): Minimizes noise estimation error when attributing intrinsic features from historical data given historical conditions.
  • Forecasting Loss ($L_F$): Minimizes noise estimation error for the future sequence, masked to supervise only the future components.
• Counterfactual Data Augmentation & Finetuning: To improve adaptability to diverse conditions without ground-truth future data, the authors construct a counterfactual dataset by randomly pairing historical sequences with alternative future conditions. They propose a finetuning algorithm that leverages semantic metrics (DTTC) to optimize the model on this synthetic counterfactual data, enhancing generalization without compromising factual accuracy.

3. Evaluation Framework (DTTC)

To evaluate forecasts in the absence of ground truth, the paper introduces DTTC (Disentangled Time Series and Text Consistency), a semantic evaluation metric based on contrastive learning.

• DTTC-I: Measures the consistency between the forecast and the intrinsic features of the historical sequence.
• DTTC-E: Measures the consistency between the forecast and the extrinsic future textual conditions.
  The DTTC model employs a dual-encoder architecture (Time Series Encoder and Text Encoder) trained to align intrinsic features across time and extrinsic features between time series and text.

Key Contributions

1. Task Formulation: Introduction of Counterfactual Time Series Forecasting with Textual Conditions, enabling flexible, condition-aware forecasting under stochastic assumptions.
2. Evaluation Framework: A comprehensive framework combining numerical accuracy (for factual settings) and semantic consistency (DTTC) for both factual and counterfactual settings, addressing the lack of ground truth in hypothetical scenarios.
3. Model Architecture: Development of TADIFF, a diffusion model featuring a text-attribution mechanism that disentangles intrinsic historical patterns from textual contexts.
4. Training Strategy: A counterfactual finetuning approach that enhances model adaptability to diverse conditions using constructed data and semantic metrics.

Experimental Results

The authors evaluated TADIFF on synthetic (Synth) and real-world datasets (ETTm1, Traffic, Exchange, Weather).

• Factual Performance: TADIFF achieved superior numerical accuracy (MAE, MSE) and semantic consistency (DTTC-I, DTTC-E) compared to unimodal baselines (e.g., DLinear, PatchTST) and multimodal baselines (e.g., VerbalTS, TimeMMD, IATSF).
• Counterfactual Performance: In settings without ground truth, TADIFF demonstrated strong generalization, achieving the highest DTTC scores, indicating it effectively balances historical constraints with hypothetical future conditions.
• Ablation Studies:
  • Text-Attribution: Removing this mechanism significantly degraded both semantic consistency and numerical accuracy, particularly in counterfactual settings, confirming its role in decoupling intrinsic features.
  • Counterfactual Finetuning: Finetuning on counterfactual data improved semantic consistency without sacrificing numerical accuracy on factual data.
  • Initial Noise: Analysis showed that text-attribution optimizes the initial noise space of the diffusion model, forming more compact clusters and leading to more effective conditional denoising.

Significance and Claims

The paper claims that TADIFF opens new avenues for modeling and assessing counterfactual forecasts under complex conditions. By moving beyond reliance on historical trajectories and simple categorical interventions, the proposed method allows for more rational and flexible forecasting in scenarios where future dynamics deviate significantly from the past.

The authors highlight three potential positive societal impacts:

1. Assisting decision-making under complex, stochastic assumptions (e.g., "What if a storm hits during a major event?").
2. Providing innovative evaluation methods for time series forecasting where ground truth is unavailable.
3. Inspiring better utilization of multimodal information in time series analysis.

The paper maintains a modest stance regarding limitations, noting challenges in evaluating forecasts of varying lengths with model-based metrics, but emphasizes the utility of the proposed framework for real-world decision-making.