When Geoffrey Négiar, PhD, talks about the origins of The Forecasting Company, the applied-AI startup he leads as co-founder and CEO, he points back to ICSI. A former graduate student researcher at ICSI, Négiar now heads a Y Combinator–backed company whose foundation models for time-series data rank among the strongest in their field, holding their own on public benchmarks against far larger systems from companies like Google, Amazon, and IBM.

Foundations at ICSI

Dr. Négiar joined ICSI in 2020 as a graduate student researcher while completing his PhD in electrical engineering and computer science at UC Berkeley, where he was part of the Berkeley Artificial Intelligence Research (BAIR) lab. Working with ICSI Vice President Dr. Michael Mahoney, he spent what he calls the “hinge period” of his research career at the institute, moving from his earlier doctoral work in mathematical optimization into deep learning.

At ICSI, his research focused on making machine-learning models more robust to adversarial attacks and on building models that can respect hard constraints such as the laws of physics, rather than merely fitting patterns in data. That dual concern, with how models behave and how to hold them to real-world rules, would become central to the company he went on to build. Dr. Négiar had come to Berkeley from France’s École Polytechnique and Ecole Normale Supérieure, and in parallel to his years at ICSI he spent time with research teams at Amazon, Google Brain, and Bloomberg before pursuing entrepreneurship.

Members of The Forecasting Company team
Members of The Forecasting Company team

Building The Forecasting Company

Dr. Négiar co-founded The Forecasting Company in 2024 with CTO Dr. Joachim Fainberg, assembling a Paris-based team of researchers and engineers drawn from labs and companies including UC Berkeley’s BAIR, MILA, Inria, JP Morgan, Google Brain, and Bloomberg. Their premise is that the same advances that produced large language models can be turned toward temporal data, such as the streams of time-stamped measurements that businesses and scientists rely on to anticipate what comes next in order to allocate resources accordingly. The company’s goal is to make accurate forecasting something teams can adopt directly, without standing up a large data-science operation or building a bespoke model for every new problem.

That work centers on the t0 family of open-weight foundation models, which forecast future values from a series’ own history together with related signals, and do so without task-specific retraining. The latest, t0-gamma, a 256-million-parameter model released in October 2026, was the #1 model on three leading public forecasting benchmarks, GIFT-Eval, fev-bench, and TIME upon release, among models evaluated without task-specific retraining. t0-gamma is a top model for power grid management, as evidenced by EnergyBench. It is a striking result for a model that is a fraction of the size of some rivals and was developed at a fraction of the cost. The company details the model in an accompanying technical report and has published its weights openly for others to test.

Members of The Forecasting Company team
Members of The Forecasting Company team

From Berkeley Research to a Growing Company

Dr. Négiar’s path reflects a pattern ICSI knows well: foundational research pursued at the institute finding its way into companies that carry those ideas out into the world. His story adds to a lineage of ICSI-connected ventures, and underscores the institute’s role as a place where early-career researchers can conduct ambitious, self-directed research and scale it to reach a global impact.

What made ICSI a good place to pursue these projects?

““My experience at ICSI with Dr. Michael Mahoney was instrumental to what I do now, and to founding The Forecasting Company. My two years at ICSI were the hinge period where I went from my earlier PhD research in optimization to really diving head first into deep learning methods. At ICSI, I worked on how to make models robust to adversarial attacks, and how to incorporate hard constraints like physics into deep learning models. At The Forecasting Company, we build foundation models for temporal data, starting with forecasting. A lot of the work we do is still inspired by ICSI and research from Berkeley.”


Geoffrey Négiar, PhD.

Former graduate student researcher at ICSI, Co-Founder at The Forecasting Company (YC S24)

This story was published in November 2026 as part of a series highlighting ICSI’s researchers and their impact. To learn about our ongoing work, explore our Core Research Themes.