Kristian Fagerlie demonstrates a developing AI forecasting system using examples from prediction marketsA prediction market lets participants trade contracts whose value depends on the outcome of a future event.. A submitted event triggers research lookups and eight parallel forecasters, after which the interface compares the resulting probabilities with the market. The demonstration exposes reasoning, sources, base-rate context and agreement between forecasters rather than presenting only a final number.
The examples show how model estimates can differ from market probabilities and how agreement can vary across questions. Kristian Fagerlie notes that one example falls outside the system's deeply tested range and that another produces low agreement. These caveats matter: disagreement with a market is not proof of an exploitable edge, and selected demonstrations or reported returns do not establish reliable out-of-sample performance.
The system remains a prototype under calibrationConfidence calibration measures whether an AI system's stated certainty matches how often its predictions or assumptions are actually correct., with custom questions and broader coverage described as future goals. The useful technical lesson is the combination of evidence gathering, multiple forecasts and inspectable uncertainty. The video does not provide enough evaluation detail to establish that the system consistently predicts events better than markets or human forecasters.
Watch on YouTube
