We rationalists are very good at making predictions, and the best of us, such as Scott Alexander

This weird self-congratulatory tribalism is a completely unnecessary distraction from an otherwise informative post. Are "we" unusually good at making predictions, compared to similarly informed and motivated people who haven't pledged fealty to our social group? How do you know?

Scott Alexander is a justly popular writer, and I've certainly benefitted from reading many of his posts, but it seems cultish and bizarre to put him on a pedestal as "the best" of "us" like this as the first sentence of a post that has nothing to do with him.

I think you're advocating two things here:

1. Make a continuous forecast when forecasting a continuous variable
2. Use a normal distribution to approximate your continuous forecast

I think that 1. is an excellent tip in general for modelling. Here is Andrew Gelman making the same point

However, I don't think it's actually always good advice when eliciting forecasts. For example, fairly often people ask whether or not they should make a question on Metaculus continuous or binary. Almost always my answer is "make it binary". Binary questions get considerable more interest and are much easier to reason about. The additional value of having a more general estimate is almost always offset by:

1. Fewer predictors => less valuable forecast
2. People update less frequently => Stale forecast
3. Harder to visualize changes over time => Less engagement from the general public

I think your point 2. has been well dealt with by gbear605, but let me add my voice to his. Normal approximations are probably especially bad for lots of things we forecast. Metaculus uses a logistic distribution by default because it automatically includes slightly heavier tails than normal distributions.

Foretold (https://www.foretold.io/) supports many continuous functions including normal for predictions and resolutions. It also had scoring rules for continuous predictions and resolution functions, and composite functions for both. The creator, Ozzie Gooen, was working on an even more sophisticated system but I'm not sure what stage that's currently at.

I generally agree with the idea - a range prediction is much better than a binary prediction - but a normal prediction is not necessarily the best. It’s simple and easy to calculate, which is great, but it doesn’t account for edge cases.

Suppose you made the prediction about Biden a couple months before the election, and then he died. If he had been taken off of the ballot, he would have received zero votes, and even if he had been left on “Biden” would have received much fewer votes. Under your normal model, the chance of either of those happening is essentially zero, but there was probably a 1-5% chance of it happening. You can adjust for this by adding multiple normal curves and giving different weights to each curve, though I’m not sure how to do the scoring with this.

It also doesn’t work well for exponential behavior. For COVID cases in a given period, a few days difference in changing behavior could alter the number of deaths by a factor of 2 or more. That can be easily rectified though by putting your predictions in log form, but you have to remember to do that.

Overall though, normal predictions work well for most predictions, and we’d be better off using them!

One of the things I like about a Brier Score is that I feel like I intuitively understand how it rewards calibration and also decisiveness.

It is trivial to be perfectly calibrated on multiple choice (with two choices being a "binary" multiple choice answer) simply by throwing decisiveness out the window: generate answers with coin flips and give confidence for all answers of 1/N.  You will come out with perfect calibration, but also the practice is pointless, which shows that we intuitively don't care only about being calibrated.

However, this trick gets a very bad (edited from low thanks to GWS for seeing the typo) Brier Score, because the Brier Score was invented partly in response to the ideas that motivate the trick :-)

We also want to see "1+1=3" and assign it "1E-7" probability, because that equation is false and the uncertainty is more likely to come from typos and model error and so on.  Giving probabilities like this will give you very very very low Brier Scores... as it should! :-)

The best possible Brier Score is 0.0 in the same way that the best RMSE is 0.0. This is reasonable because the RMSE and Brier Score are in some sense the same concept.

It makes sense to me that for both your goal is to make them zero. Just zero. The goal then is to know all the things... and to know that you know them by getting away with assigning everything very very high or very very low probabilities (and thus maxing the decisiveness)! <3

Second we calculate ${\hat{\sigma }}_z$ as the RMSE (root mean squared error) of all predictions... Then we calculate ${\hat{\sigma }}_z$ ...

$$...=1.73$$

So if these were my only two predictions, then I should widen my future intervals by 73%. In other words, because ${\hat{\sigma }}_z$ is 1.73 and not 1, thus my intervals are too small by a factor of 1.73.

I'm not sure if you're doing something conceptually interesting here (like how Brier Scores interestingly goes over and above mere "Accuracy" or mere "Calibration" by taking several good things into account in a balanced way), or... maybe... are you making some sort of error? 

RMSE works with nothing but point predictions. It seems like you recognize that the standard deviations aren't completely necessary when you write:

(1) If the data $x$ and the prediction $\mu$ are close, then you are a good predictor

Thus maybe you don't need to also elicit a distribution and a variance estimate from the predicter? I think? There does seem to be something vaguely pleasing about aiming for an RMSE of 1.0 I guess (instead of aiming for 0.00000001) because it does seem like it would be nice for a "prediction consumer" to get error bars as part of what the predictor provides?

But I feel like this might be conceptually akin to sacrificing some of your decisiveness on the altar of calibration (as with guessing randomly over outcomes and always using a probability of 1/N).

The crux might be something like a third thing over and above "decisiveness & calibration" that is also good and might be named... uh... "non-hubris"? Maybe "intervalic coherence"? Maybe "predictive methodical self-awareness"?

Is it your intention to advocate aiming for RMSE=1.0 and also simultaneously advocate for eliciting some third virtuous quality from forecasters?

(If this makes no sense, then ignore it): Using an arbitrary distribution for predictions, then use its CDF (Universality of the Uniform) to convert to $U(0,1)$, and then transform to z-score using the inverse CDF (percentile point function) of the Unit Normal. Finally use this as $z_i$ in when calculating your calibration.

Well, this makes some sense, but it would make even more sense to do only half of it.

Take your forecast, calculate it's percentile. Then you can do all the traditional calibration stuff. All this stuff with z-scores is needlessly complicated. (This is how Metaculus does it's calibration for continuous forecasts)

Curated. Predictions are the bread and butter of rationality, so I love this post advising on how to systematically do them better.

It is a pretty big ask of individuals (who perhaps are making a blog post with a list of yearly predictions, in the style of Slate Star Codex, theZvi, Matt Yglesias, or others) to do all this math in order to generate and evaluate Normal Predictions of continuous variables. I think your post almost makes more sense as a software feature request -- more prediction markets and other platforms should offer Metaculus-style tools for tweaking a distribution as a way of helping people generate a prediction:

• It would be awesome if a prediction market let me bet on the 2024 election by first giving me a little interactive tool where I could see how my ideas about the popular vote probability distribution might translate into a Democratic or Republican victory. Then I could bet on the binary-outcome market after exploring my beliefs in continuous-variable land.
• Perhaps blogging platforms like LessWrong or Substack could have official support for predictions -- maybe you start by inputting your percentage chance for a binary outcome, and then the service nudges you about whether you'd like to turn that prediction into a continuous variable, to disambiguate cases like your example of "N(50.67, 0.5), N(54, 3), N(58, 6)" all giving 91% odds of a win. This seems like a good way of introducing people to the unfamiliar new format of Normal Predictions.

Thanks, I really enjoyed this post - this was a novel but persuasive argument for not using binary predictions, and I now feel excited to try it out!

One quibble - When you discuss calculating your calibration, doesn't this implicitly assume that your mean was accurate? If my mean is very off but my standard deviation is correct, then this method says my standard deviation is way too low. But maybe this is fine because if I have a history of getting the mean wrong I should have a wider distribution?

Any thoughts on mean deviation? Seems more intuitive in some ways

(If this makes no sense, then ignore it): Using an arbitrary distribution for predictions, then use its CDF (Universality of the Uniform) to convert to $U(0,1)$, and then transform to z-score using the inverse CDF (percentile point function) of the Unit Normal. Finally use this as $z_i$ in when calculating your calibration.

This glosses over an important issue: namely, how do you find (and score once you do find it) which distribution is correct?

(This issue occurs surprisingly often in my experience - one person thinks that a distribution has a thin tail, and another thinks that the distribution has a fat tail, but there's little enough data that both estimates are plausible. Of course, they give very different predictions for long-tail events.)

(Also, P(0 votes for Biden) is arguably significantly more likely than P(1 vote for Biden), in which case even the example chosen suffers from this issue of choosing the wrong distribution.)