I really liked this post when I read it, even though I didn’t quite understand it to my satisfaction (and still don’t).
As far as I understand the post, it proposes a concrete & sometimes fleshed-out solution for the perennial problem of asking forecasting questions: How do I ask a forecasting question that actually resolves the way I want it to, in retrospective, and doesn’t just fall on technicalities?
The proposed solution is latent variables and creating prediction markets on them: The change in the world (e.g. AI learning to code) is going to affect a bunch of different variables, so we set up a prediction market for each of those, and then (through some mildly complicated math) back out the actual change in the latent variable, not just the technicalities on each of the markets.
This is pretty damn great: It’d allow us to more quickly communicate changes in predictions about the world to people and read prediction market changes more easily. The downside is that betting on such markets would be more complicated and effortful, and especially for non-subsidized markets that is often a breaking point. (Many “boring” Metaculus questions receive <a few dozen predictions.)
I’m pretty impressed by this work, and wish there was some efforts at integration into existing platforms. Props to tailcalled for writing code and an explanation for this post. I’ll give it +4, maybe even +9, because it represents genuine progress on a tricky intellectual question.
I really liked this post when I read it, even though I didn’t quite understand it to my satisfaction (and still don’t).
As far as I understand the post, it proposes a concrete & sometimes fleshed-out solution for the perennial problem of asking forecasting questions: How do I ask a forecasting question that actually resolves the way I want it to, in retrospective, and doesn’t just fall on technicalities?
The proposed solution is latent variables and creating prediction markets on them: The change in the world (e.g. AI learning to code) is going to affect a bunch of different variables, so we set up a prediction market for each of those, and then (through some mildly complicated math) back out the actual change in the latent variable, not just the technicalities on each of the markets.
This is pretty damn great: It’d allow us to more quickly communicate changes in predictions about the world to people and read prediction market changes more easily. The downside is that betting on such markets would be more complicated and effortful, and especially for non-subsidized markets that is often a breaking point. (Many “boring” Metaculus questions receive <a few dozen predictions.)
I’m pretty impressed by this work, and wish there was some efforts at integration into existing platforms. Props to tailcalled for writing code and an explanation for this post. I’ll give it +4, maybe even +9, because it represents genuine progress on a tricky intellectual question.