Introduction

https://astralcodexten.substack.com/p/mantic-monday-1115

I talked about this last week as a potential solution to the problem of long-term forecasting. Instead of waiting a century to see what happens, get a bunch of teams, and incentivize each to predict what the others will guess. If they all expect the others to strive for accuracy, then the stable Schelling point is the most accurate answer.

Now there’s a paper, by Karger, Monrad, Mellers, and Tetlock - Reciprocal Scoring: A Method For Forecasting Unanswerable Questions.

They focus not just on long-run outcomes but on conditionals and counterfactuals. The paper starts with an argument against conditional prediction markets that I’d somehow missed before. Suppose you want to know whether a mask mandate will save lives during a pandemic. Current state of the art is to start two prediction markets: “conditional on there being a mask mandate, how many people will die?” and “conditional on there not being a mask mandate, how many people will die?” In this situation, this doesn’t work! Governments are more likely to resort to mask mandates in worlds where the pandemic is very bad. So you should probably predict a higher number of deaths for the mandate condition. But then confused policy-makers will interpret your prediction market as evidence that a mask mandate will cost lives.