Can’t think of anything better than a t-test, but open for suggestions.
If a forecaster is consistently off by like 10 percentage points—I think that is a difference that matters. But even in that extreme scenario where the (simulated) difference between two forecasters is in fact quite large, we have a hard time picking that up using standard significance tests.
Interesting, thanks for sharing the paper. Yeah agree that using the Brier score / log score might change results and it would definitely be good to check that as well.
In principle yes. In practice also usually yes, but the specifics depend on whether the average user who predicted on a question gets a positive amount of points. So if you predicted very late and your points are close to zero, but the mean number of points forecasters on that question received is positive, then you will end up with a negative update to your reputation score.
Completely agree that a lot hinges on that reputation score. It seems to work decent for the Metaculus Prediction, but it would be good to see what results look like for a different metric of past performance.
Not sure how to quantify that (open for ideas). But intuitively I agree with you and would suspect it’s at least a sizable part
Yeah, definitely. The title was a bit tongue-in-cheek (it’s a movie quote)
And is the code to the MetaculusBot public somewhere? :)
It should be possible to fully automate the bot and just run a CRON job that regularly checks the Metaculus API for new questions, right?
I slightly tend towards yes, but that’s mere intuition. As someone on Twitter put it, “Metaculus has a more hardcore user base, because it’s less fun”—I find it plausible that the Metaculus user base and the Manifold user base differs. But higher trading volume I think would have helped.
For this particular analysis I’m not sure correcting for the number of forecasters would really be possible in a sound way. It would be great to get the MetaculusBot more active again to collect more data.
Is it possible to get rid of the question mode for this post?
For Metaculus there are lots of ways to drive engagement: prioritise making the platform easier to use, increase cash prizes, community building and outreach etc.
But as mentioned in the article the problem in practice is that the bootstrap answer is probably misleading, as increasing the number of forecasters likely changes forecaster composition.
However, one specific example where the analysis might be actually applicable is when you’re thinking about how many Pro Forecasters you hire for a job.
In principle yes, you’ll just still always have the problem that people are predicting at different time points. If the best and the 2nd best predict weeks or months apart then that changes results.
Ah snap! I forgot to remove that paragraph… I did subsampling initially, then switched to bootstrapipng. Resulsts remained virtually unchanged. Thanks for pointing that out, will update the text.
Hi Simon, I’m working on a follow-up to this post that uses individual-level data. Could you please give some detail on how you “sampled” k predictors? As in, did you have access to individual data and could actually do the sampling? I’m not entirely sure what the x-axis in your plot means and what the difference betwenn “>N predictors” and “k predictors” is. Thank you!
Good comment, thank you!