Truthful Calibration Measures for Sequential Prediction
Exact truthfulness cannot coexist with sound and complete calibration scoring for sequential binary forecasts.
The paper proves that negative result even when outcomes are independent. It then narrows the failure to exact truthfulness, giving two reductions that turn a base calibration measure into approximately truthful ones. One construction yields a sound and complete measure with a multiplicative truthfulness guarantee of `1 + exp(-T^((1-epsilon)/2)/2)`, improving Haghtalab et al. (2024). ArXiv · AI/CL/LG's note
The paper proves that negative result even when outcomes are independent. It then narrows the failure to exact truthfulness, giving two reductions that turn a base calibration measure into approximately truthful ones. One construction yields a sound and complete measure with a multiplicative truthfulness guarantee of `1 + exp(-T^((1-epsilon)/2)/2)`, improving Haghtalab et al. (2024). ArXiv · AI/CL/LG's note
score 4