Abstract
The quality of probabilistic forecasts is crucial for decision-making underuncertainty. While proper scoring rules incentivize truthful reporting ofprecise forecasts, they fall short when forecasters face epistemic uncertaintyabout their beliefs, limiting their use in safety-critical domains wheredecision-makers (DMs) prioritize proper uncertainty management. To addressthis, we propose a framework for scoring imprecise forecasts -- forecasts givenas a set of beliefs. Despite existing impossibility results for deterministicscoring rules, we enable truthful elicitation by drawing connection to socialchoice theory and introducing a two-way communication framework where DMs firstshare their aggregation rules (e.g., averaging or min-max) used in downstreamdecisions for resolving forecast ambiguity. This, in turn, helps forecastersresolve indecision during elicitation. We further show that truthfulelicitation of imprecise forecasts is achievable using proper scoring rulesrandomized over the aggregation procedure. Our approach allows DM to elicit andintegrate the forecaster's epistemic uncertainty into their decision-makingprocess, thus improving credibility.