Alignment Problems With Current Forecasting Platforms
Abstract
We present alignment problems in current forecasting platforms, such as Good Judgment Open, CSET-Foretell or Metaculus. We classify those problems as either reward specification problems or principal-agent problems, and we propose solutions. For instance, the scoring rule used by Good Judgment Open is not proper, and Metaculus tournaments disincentivize sharing information and incentivize distorting one's true probabilities to maximize the chances of placing in the top few positions which earn a monetary reward. We also point out some partial similarities between the problem of aligning forecasters and the problem of aligning artificial intelligence systems.
Keywords
Cite
@article{arxiv.2106.11248,
title = {Alignment Problems With Current Forecasting Platforms},
author = {Nuño Sempere and Alex Lawsen},
journal= {arXiv preprint arXiv:2106.11248},
year = {2023}
}
Comments
39 pages, 13 figures. v2: revised after email with Good Judgment Inc about IARPA's ACE tournament--it did use an improper scoring rule, but they claim that the effect was quantitatively small. Revised v3: fix typos, finish unfinished sentence