Zoa Research

Zoa Research is developing powerful quantitative forecasting models that transcend the limitations of domain-specific approaches. Traditionally, the creation of such models has required extensive, domain-specific expertise, with experts dedicating years to feature engineering, hyperparameter tuning, and architectural iteration within narrow fields. Zoa Research believes that scale is the key to unlocking superior forecasting capabilities. By training cross-domain event forecasting engines, they aim to uncover patterns that specialized models and human intuition might miss.

Their approach involves automating the iterative modeling cycle using LLMs embedded in multi-agent optimization loops, evaluated against fixed policies. This allows for continuous improvement, similar to AlphaEvolve, but tailored for forecasting problems. Unlike existing models that rely heavily on human priors, Zoa Research leverages data from diverse contexts to build sample-efficient general models. These models incorporate more inductive priors and rely more heavily on inference-time compute compared to LLMs, enhancing sample efficiency.

The potential applications of improved forecasting are vast, spanning supply chain management, energy sector predictions, and even natural disaster risk assessment. By enhancing our ability to predict future events, Zoa Research aims to improve decision-making across various industries and scientific domains. Their models can be used by labs and academics in data-heavy domains, ultimately contributing to scientific progress and a better understanding of the world.