Neurofusion says its forecasting model topped ForecastBench's leaderboard, ahead of Google DeepMind
The AMW Read
A Korean fintech AI startup claims a self-reported preliminary benchmark win over Google DeepMind and xAI on future-event forecasting, a notable but not yet independently verified sub-segment capability signal.
Neurofusion says its forecasting model topped ForecastBench's leaderboard, ahead of Google DeepMind
Neurofusion, the South Korean company behind investment platform Valley AI, said its research team placed first on the preliminary leaderboard of ForecastBench, which scores AI systems on unresolved economic, financial, political and technology events using data from FRED, Yahoo Finance, and prediction markets Polymarket and Metaculus. Run by the Forecasting Research Institute, whose contributors include "Superforecasting" author Philip Tetlock, the benchmark refreshes questions every two weeks. Neurofusion's model scored 67.5 of 464 submissions, ahead of the fifth-place team (65.9), Google DeepMind's top entry (65.0), and the 63.6 average of the benchmark's human "superforecaster" panel. A full combined ranking awaits ForecastBench's separate market-prediction track, where xAI also competes.
The score highlights specialized architecture competing with frontier general models on a narrow, commercially loaded task: predicting market and economic outcomes ahead of resolution. Neurofusion attributes it to Demian, a time-series graph structure built over four years across more than 530 data pipelines, arguing forecasting needs quant and general LLM reasoning plus a separate model of how human market participants think and feel — with the company saying only about half of that architecture exists today.
Neurofusion, founded by former S&P Global data scientist and hedge-fund trader Choi Han-cheol, plans to fold the forecasting capability into Valley AI's research function alongside CereFin, a financial-reasoning agent added to the platform this month. Valley AI has drawn 10,000 paid subscribers and more than 30,000 free members since its January launch. For finance-AI builders and investors, the case for a differentiated forecasting layer over a general-model wrapper strengthens, though the claim is self-reported and the underlying architecture is admittedly unfinished.
