Transparency in Forecasting

Our Experimental Framework

Our methodology uses questions / framework from ForecastBench.org. We ensure fair comparison between human and AI forecasters through careful experimental controls.

• Data Leakage Prevention: We strictly limit data leakage by using models with training cutoffs before forecast dates, and only allow access to articles last updated before the forecast date to supplement knowledge.
• Statistical Significance: All results are tested for statistical significance using both bootstrap methods and Mann-Whitney U tests to ensure robust conclusions.