Related: It would be cool to have a Slack or Discord channel for chatting about Metaculus and Metaculus-adjacent topics.

It's a more productive-feeling procrastination outlet than anything else I've found on the internet.

The value on February 27 was ~0.087. After that, the WHO changed the format of the report, no longer breaking out cases by origin in China, and instead reporting whether cases are believed to represent imported or community infections. Accordingly, the Feb 27 number is very likely to be the last time this can be computed, and according to the question text, the number which should be the resolution.

Show Brier and log scores on resolved question pages. I know we can see them on our personal track records, but it would be nice to see them in the resolved question pages next to the points won.

Also, bigger lift: how you did relative to the community. I'm thinking something like "Out of n forecasters, your prediction was more accurate than x of them. You beat the community median score by y points."


I think other questions like that would be interesting: i.e., P(some stuff| "superintelligent AGI")

Oh, we should totally do a "Seeing Past the Singularity" series!…

It appears Oman was not removed from the list as of 2019-10-10, and a quick perusal of the other press releases suggests that re-evaluations don't occur routinely enough to make it especially likely Oman will be removed prior to the deadline.

Polynomial fit predicts ~18k cases [on Feb 3]( [Extended outwards to Feb 17, predicts ~85k cases]( Set wide margins because: * The polynomial fit is good, but it's a weak model--theory suggests exponential fit should more closely represent act...

One useful note is that there are only 11 Democratic Incumbents running. Accordingly, predictors should optimize their scores by creating minimum variance spikes on multiples of 9 and setting their weights according to your estimates of the likelihood of the corresponding outcome.

(The others in this set are somewhat less vulnerable to this approach to score-hacking.)

@Based No, the optimal strategy is to forecast your actual credence in the hypothesis that the question will resolve positively. Some will resolve early leading to comparatively tiny scores, but if you try to game the system on that information, you'll end up losing points long term.

I'm inclined to say it should resolve negatively. After all, if there is no Mars, there definitely aren't 10E6 people living on it.

Someone's probably already pitched this, but... New Question type: Vector. Instead of giving a distribution over a range of possibilities, give an array of point predictions, and then resolve them as they become available. For example, Predict the daily close value of the S&P500 every day for the next month. Presently, Metaculus can do this with 31 discrete questions, one for each day. But it would be cool if that could be a single question, into which all 31 predictions are entered. (Similar to a Kaggle competition.)

@notany The Bloomberg data isn't being maintained AFAICT. I've been tracking Musk predictions here. His on-time delivery record is worse than you think.

Tachyon Tipping: Sometimes somebody posts an update that I get a lot of benefit from and I'd like to tip them, somehow. It'd be cool if I could send them a tachyon.

@AABoyles Here's a colab notebook leveraging @ought's excellent ergo library to normalize the community estimates into consistent predictions.

There are a bunch of things Metaculus doesn't do that I want it to. Some of these can be accomplished with just a little bit of javascript. So I created this:

I don't have any tracking on the site, so please let me know if you get any value from it.