We didn't just top ForecastBench. We swept it. On our first submission, the three best-performing models on the ForecastBench preliminary dataset leaderboard are all Torchcast AI. Not just #1. The whole podium. 🥇🥈🥉
ForecastBench, designed by the Forecasting Research Institute, scores AI systems on real, unresolved future events: markets, macro, geopolitics, and the questions where uncertainty actually costs you. The field includes the strongest AI labs and forecasting systems in the world, benchmarked against the median of a group of human superforecasters.
Three Torchcast systems now sit at the top: ahead of all frontier labs and human superforecasters.
This is not just a bigger model or a better prompt. It is calibration. We trained Torchcast to know what it knows; and just as importantly, what it does not. A useful forecast is not the most confident story. It is a probability you can act on. If a system says 70%, it should be right about 70% of the time. That sounds simple. It is extremely hard.
Most AI is built to sound confident. We are building systems that reason under uncertainty, update with evidence, and produce probabilities that institutions can use to make high-stakes decisions.
For a young startup that started just two months ago, this is a meaningful signal. It tells us we are working on the right problem, and reinforces our belief that forecasting may become one of the most important measures of real intelligence in AI.
I'm incredibly proud of our small team of three, He Li and Zhilin Hu, and grateful to everyone behind Torchcast AI — our investors, advisors, customers, and supporters. This is what disciplined, unglamorous work looks like when it compounds. We are just getting started.
Torchcast AI works with hedge funds, commodities desks, manufacturers, institutional exporters, real estate teams, and organizations across sectors that would rather price the future than guess at it. If that is you, my inbox is always open.
See the full ForecastBench preliminary dataset leaderboard ↗