Georg M. Goerg

Research

Statistical AI research on causality, forecasting and world models, built into products that work.

Since I joined EvolutionIQ in March 2020, most of my time has gone into building products and growing the company, so I’ve published less research. The work didn’t stop, though: the 2022 CLeaR paper on PSPS, a patent application at EvolutionIQ, and most recently the 2026 NeurIPS workshop paper with Igor Shuryak.

Setting a Dose versus Observing One: Continuous Dose-Response Estimation under Right-Censoring Georg M. Goerg and Igor Shuryak (Columbia University). Accepted at the NeurIPS 2026 Workshop on World Models for High-Stakes Health (WMHS).

PSPS-CS is a deep mixture-of-experts model that estimates whole dose-response curves for continuous treatments, such as a radiotherapy dose, when outcomes are right-censored survival times and confounding is strong. In simulations it was competitive with five other estimators. On a real head and neck cancer cohort the data support no dose contrast and no change in current clinical practice; the model is meant as a complementary tool for hypothesis generation.

Because it estimates what happens when you set a quantity rather than just observe it, the same architecture can be the basis for world models for agents that take actions and estimate their outcomes.

OpenReview · Code: github.com/gmgeorg/pypsps

Selected publications

  • Predictive State Propensity Subclassification (PSPS): A causal inference algorithm for data-driven propensity score stratification. Joseph Kelly, Jing Kong, and Georg M. Goerg. Conference on Causal Learning and Reasoning (CLeaR), 2022. PMLR · Code: Python
  • Predictive State Smoothing (PRESS): Scalable non-parametric regression for high-dimensional data with variable selection. Georg M. Goerg. Google technical report, 2017. Google Research · Code: Python
  • The Lambert Way to Gaussianize Heavy-Tailed Data with the Inverse of Tukey’s h Transformation as a Special Case. The Scientific World Journal, 2014. Code: R, Python, PyTorch
  • Forecastable Component Analysis. ICML 2013. Code: R, Python
  • Lambert W random variables: a new family of generalized skewed distributions with applications to risk estimation. Annals of Applied Statistics, 2011. Code: R, Python, PyTorch

The full list is on Google Scholar and my Google research profile.

Patents

  • Systems and methods of anomaly detection. Farhana Bandukwala, Georg M. Goerg, Samuel Paul Ruth, and Jonathan Jesse Halcrow; assigned to Google LLC. US 12,353,316 B2, granted July 8, 2025 (published as US 2023/0138371 A1). Google Patents

I’m also an inventor on another USPTO application that was allowed and then withdrawn before issue, so its content isn’t public.

Thesis

  • PhD, Carnegie Mellon University (2012). Learning Spatio-Temporal Dynamics: Nonparametric Methods for Optimal Forecasting and Automated Pattern Discovery. Advisors: Cosma Shalizi and Larry Wasserman.
  • Diplom-Ingenieur, Vienna University of Technology (2007). Long Memory versus Structural Breaks: A Time-Varying Memory Approach. Advisor: Manfred Deistler.

Software

Open-source packages are listed on the Software page.