Teaching Models to Imagine: Trustworthy Simulation for Sequential Data

When real-world sequential data is sparse, biased, delayed, or impossible to access, how do we train models to reason well instead of hallucinating patterns? This talk will explain how to generate synthetic and simulated sequential data in ways that are distribution-agnostic, robust, and trustworthy enough to support modern ML systems, including transformer-based models. The talk will show how simulation can be used not just for prediction, but also for recommendations, policy evaluation, and counterfactual reasoning in dynamic environments. Attendees will learn practical principles for designing simulators that preserve the right structure, stress-test models under shift, and avoid misleading conclusions. The session will be especially useful for practitioners who need better data-generation strategies when observational data is incomplete or operationally constrained, and want a more reliable path from limited real-world data to models that generalize, adapt, and support better decisions.