Intervention-Aware Forecasting: Breaking Historical Limits from a System Perspective

TL;DR

This work reframes time series forecasting as dynamic system modeling by explicitly incorporating external interventions, with a focus on textual signals. It provides a theoretical basis that ignoring interventions creates an irreducible error floor and demonstrates that including interventions can lower error bounds. The FIATS model, featuring CASM and CAPS, and the leak-free Temporal-Synced IATSF benchmark show that intervention-aware modeling yields consistent, significant gains across synthetic, physical, and market domains, outperforming state-of-the-art baselines—including LLM-assisted methods—without relying on increased architectural complexity. The results highlight the practical value of textual cues as causal intervention signals and offer interpretable mechanisms to understand channel-specific sensitivities and causal alignments. This framework paves the way for more robust, adaptable forecasting in real-world systems where external interventions are pervasive and qualitatively described.

Abstract

Traditional time series forecasting methods predominantly rely on historical data patterns, neglecting external interventions that significantly shape future dynamics. Through control-theoretic analysis, we show that the implicit "self-stimulation" assumption limits the accuracy of these forecasts. To overcome this limitation, we propose an Intervention-Aware Time Series Forecasting (IATSF) framework explicitly designed to incorporate external interventions. We particularly emphasize textual interventions due to their unique capability to represent qualitative or uncertain influences inadequately captured by conventional exogenous variables. We propose a leak-free benchmark composed of temporally synchronized textual intervention data across synthetic and real-world scenarios. To rigorously evaluate IATSF, we develop FIATS, a lightweight forecasting model that integrates textual interventions through Channel-Aware Adaptive Sensitivity Modeling (CASM) and Channel-Aware Parameter Sharing (CAPS) mechanisms, enabling the model to adjust its sensitivity to interventions and historical data in a channel-specific manner. Extensive empirical evaluations confirm that FIATS surpasses state-of-the-art methods, highlighting that forecasting improvements stem explicitly from modeling external interventions rather than increased model complexity alone.

Figures (14)

• Figure 1: The real system runs under various interventions. The Intervention-Aware method can effectively approximate the real system according to the dataset while traditional self-stimulated method can only approximate a average scenario with persistent error, lead to bad or even collapse result. The right panel shows visualization result of a frequency modulated system which is very sensitive to the intervention, i.e. large $\nabla_U F$.
• Figure 2: Architecture of FIATS. FIATS integrates three inputs: time series data from a look-back window, temporal-synced news embeddings, and channel description embeddings. The intervention encoder employs CASM blocks in a residual connection along with multiple self-attention layers to enhance feature extraction. The CAPS causal alignment decoder projects the historical time series embeddings into the future, guided by channel-aware, time-synced interventions. A token-wise decoder is used to prevent overfitting in the final linear layer, as discussed in lee2023learning.
• Figure 3: Performance improvement with respect to the PatchTST on each time series in GAUD.
• Figure 4: Visualization of three channels on the 15,000th test sample of the Atmos. Phy. 2014-19 dataset. Blue indicates ground truth, Red shows FIATS, Green represents PatchTST, and Orange denotes FIATS with swapped interventions on the second and fourth forecast days. The CAPS causal alignment decoder exhibits distinct attention patterns across channels.
• Figure 5: Attention map of the CASM on the 15000th test sample of Atmos. Phy. 2014-19 dataset. We use three cross attention block in residual connection. The horizontal axis stands for channels and vertical stands for the 7 sentences of the weather report summary.

Theorems & Definitions (8)

• Proposition 2.1: Self-Stimulation Error Bound
• Proposition 3.1: Partial Intervention Efficacy
• proof
• proof
• proof
• Proposition B.1: Self-Stimulation Error Floor
• proof : Justification of Proposition \ref{['thm:error_bound_general']}
• proof