time series and causal inference

This study addresses the problem of prospective causal forecasting in panel data—specifically, predicting counterfactual outcomes for a target unit over future periods during which an intervention has not yet been implemented. To this end, the authors propose the Two-Way Synthetic Forecasting (TWSF) estimator, which uniquely integrates synthetic control methods with multivariate time series extrapolation. The approach leverages a low-rank latent factor model to capture both cross-sectional dependencies and temporal dynamics, combining unit-specific regressors, time factor modeling, and orthogonalization-based bias correction. Theoretical analysis establishes pointwise consistency, asymptotic normality, and finite-sample error bounds for multi-step-ahead predictions. Extensive simulations confirm the estimator’s empirical performance, and a real-world application demonstrates its utility in evaluating the causal impact of 2020 NFL stadium reopenings on public health outcomes.

This paper addresses counterfactual prediction in panel data with missing observations and time-varying latent factors. We propose FOCUS—a unified framework integrating matrix completion with autoregressive modeling of latent factors, explicitly distinguishing stochastic and deterministic temporal dynamics. Built upon a PCA-based estimator, FOCUS supports finite-sample error bounds and asymptotic normality under stationarity assumptions. Theoretically and empirically, FOCUS outperforms existing benchmarks—particularly in settings with autoregressive latent factors. In application to the HeartSteps mobile health study, FOCUS accurately predicts individual-level step-count trajectories under interventions, demonstrating its validity and practical utility for real-world causal inference tasks.

This paper addresses the “spurious synthetic control problem” in macroeconomic applications of the Synthetic Control Method (SCM), wherein imposing common nonstationary trends across units induces bias in causal estimates. To resolve this, we propose the Synthetic Commercial Cycle Framework (SCCF), which adopts a “divide-and-conquer” strategy: it models the long-run trend solely using the treated unit’s own historical data, while leveraging only control units to estimate cyclical fluctuations—thereby decoupling trend and cycle components. SCCF integrates time-series decomposition with nonstationary counterfactual modeling. Applied to two canonical macroeconomic cases—German reunification and Hong Kong’s return to China—the framework substantially improves counterfactual prediction accuracy, eliminates trend misspecification bias, and enhances the robustness and reliability of causal inference for macroeconomic policy evaluation.

Traditional time-series forecasting methods neglect the causal impact of external interventions—such as news articles or policy announcements—on dynamic system evolution, leading to suboptimal accuracy. This paper proposes the Intervention-Aware Time-Series Forecasting (IATSF) framework, the first to systematically model the causal effects of textual interventions. Our contributions are threefold: (1) we construct an information-leakage-free textual intervention benchmark; (2) we introduce Channel-Adaptive Sensitivity Modeling (CASM) and Channel-Adaptive Parameter Sharing (CAPS) to capture heterogeneous intervention effects across time-series channels; and (3) we design FIATS, a lightweight model integrating control-theoretic analysis with efficient textual encoding. Evaluated on both synthetic and real-world datasets, IATSF consistently outperforms state-of-the-art methods. Crucially, ablation studies confirm that explicit intervention modeling—not increased model complexity—is the primary driver of performance gains.

Modeling dynamic, non-stationary causal structures in complex systems remains challenging due to evolving temporal dependencies and heterogeneous causal mechanisms. Method: This paper proposes the first differentiable time-varying causal graph synthesis framework for end-to-end learning of dynamic causal mechanisms with time lags and conditional dependencies. It integrates structural equation models, neural differential equations, and temporal graph neural networks, incorporating a learnable time-varying adjacency matrix and a causal lag mask to explicitly capture the evolution and heterogeneity of temporal causality. Contribution/Results: The method achieves a 12.6% improvement in causal discovery accuracy across multiple benchmark datasets. Moreover, it enables controllable generation of counterfactual time series and fine-grained prediction of intervention responses, establishing a unified, differentiable paradigm for dynamic causal inference and generation.

This study addresses the opacity of causal mechanisms and the lack of counterfactual explanations in multivariate time series forecasting. To this end, it proposes a novel framework that integrates genetic algorithms with Granger causality testing, augmented by quantile regression to model the conditional distribution under interventions. This approach uniquely combines evolutionary search with rigorous causal inference to automatically generate statistically significant and interpretable counterfactual scenarios. Experiments on real-world datasets demonstrate that the proposed method not only effectively uncovers the underlying causal structures among complex variables but also enables reliable reasoning about the outcomes of hypothetical interventions, thereby providing actionable counterfactual support for decision-making.

Existing time series benchmarks lack interventional data, hindering the training of causal foundation models. To address this gap, this work proposes CausalTimePrior, a framework that introduces the first synthetic time series structural causal model (TSCM) capable of generating paired observational and interventional data. The framework supports configurable causal graphs, nonlinear autoregressive mechanisms, state-switching dynamics, and diverse intervention types—including hard, soft, and time-varying interventions. Prior-data fitting networks (PFNs) trained within this framework demonstrate effective in-context estimation of causal effects on unseen TSCMs, underscoring the framework’s pivotal role in advancing causal foundation models for time series.

This study addresses a critical limitation in existing synthetic control methods—their neglect of temporal dynamics, which hinders effective causal inference in the presence of strong time trends. To overcome this, we propose a time-aware synthetic control framework that explicitly models temporal dynamics within the synthetic control paradigm. Our approach integrates constant trends and low-rank signal structures through a state-space model and employs Kalman filtering combined with Rauch–Tung–Striebel smoothing for counterfactual estimation. The method substantially improves the accuracy of causal effect estimation under noisy conditions. Extensive experiments on both simulated data and real-world applications—including policy evaluation and sports forecasting—demonstrate that our approach consistently outperforms current baselines, particularly when strong temporal trends and high noise levels are present.

Traditional causal inference methods struggle to capture how interventions affect the dynamic evolution of time series, such as persistence and transition patterns. This work extends the potential outcomes framework to path space and introduces the Dynamic Average Treatment Effect (DATE) to characterize how causal effects evolve over time. It establishes the first dynamic causal inference framework in path space, develops a dynamic inverse probability weighting estimator suitable for observational data, and reveals that, under sparse treatment regimes, the conditional mean trajectory admits a linear state-space structure. Simulations demonstrate that the proposed method accurately captures dynamic effects that static approaches systematically misestimate. In an empirical application to COVID-19 lockdown policies, the method successfully estimates and decomposes the treatment effects over time.

This work addresses the scarcity of ground-truth causal annotations and limited physiological data in multivariate time series (MTS) research. To overcome these challenges, we propose a practical spatiotemporal causal graph modeling framework grounded in discrete-time structural causal processes. The framework supports lagged causality, mixed variable types (continuous and discrete), and interpretable causal interventions under user-specified distribution shifts. It integrates domain expert knowledge with algorithmic suggestions to enable programmable causal graph construction and interactive modeling workflows. Edge functions are implemented via a hybrid design combining neural networks with parameterized templates, explicitly encoding causal dynamics. The resulting synthetic MTS datasets feature known, intervenable causal structures—significantly augmenting real-world datasets and providing a configurable, reproducible benchmark for evaluating causal discovery algorithms.