Vladyslav Moroshan
This PhD program proposes the development of Causal Foundation Models for Amortized Interventional Forecasting. By merging rigorous causal identifiability frameworks with the scaling capabilities of deep learning, this research amortizes the functional interventional response directly into the weights of a pre-trained sequence model. Rather than extracting an explicit directed graph at test time, the model is pre-trained on large-scale procedural data engines governed by diverse Structural Causal Model (SCM) priors. Given an observational context window, it produces a calibrated probabilistic distribution over post-intervention outcomes in a single forward pass, explicitly representing the epistemic uncertainty inherent in causal identification.
The methodology is deeply anchored in continuous-time multivariate time series. Unlike static causal models, interventional forecasting in dynamical systems must account for lagged causal cascades, high-dimensional state spaces, and irregularly sampled observations. The target estimand is the Conditional Average Potential Outcome (CAPO) trajectory, a full temporal forecast of the system's evolution following a surgical structural intervention, rather than a static scalar effect. While structurally optimized for continuous-time sequences, the underlying mathematical framework is designed to be modality-agnostic, enabling future extensions to tabular and multimodal inference.
The research is structured around four pillars: (1) a synthetic data engine generating paired factual and counterfactual multivariate SCM environments via nonlinear regime-switching stochastic differential equations; (2) the design of amortized foundation architecturesn coupled with multimodal probabilistic output heads to capture structural ambiguity without posterior mode collapse; (3) the empirical characterization of degradation limits, formally isolating dynamical fidelity loss (reconstruction error due to discrete undersampling) from causal identifiability bounds (structural equivalence classes) to map the "Sim-to-Real" gap; and (4) applied validation across physical ground-truth benchmarks (Causal Chambers), clinical target trial emulation on irregular physiological trajectories (MIMIC-IV), and econometric dynamic pricing.
Ultimately, this work aims to deliver zero-shot, interventionally coherent forecasting systems capable of rigorous uncertainty quantification for complex sequential decision-making.
