I had this question at 40–50 % when the community was at 60–75 %. People seemed to be banking on Apple following suite in line with other large technology companies, but I had a vague hunch that if there is one technology company that does not follow in the footsteps of others, it is Apple.

I stuck to 63 % for the duration of the question, as the community wavered from a high of 60 % down to 30–40 %. My reasoning was mainly that (a) it is very easy for this to resolve yes, since it only takes one mention, and (b) it might be in her interest to mention ai for signaling value, to appeal to younger voters.

I stuck to 25 % long after the community went up to 50 % and even 80 %. I just had trouble believing a candidate would quit so late in the race – even when it would be in the party’s best interest. I have to admit I don’t understand the U.S. political system. I mean, sure, it would make sense for Biden to quit. But look at the other candidate! He is also spouting nonsense, but additionally faces potential jail time for quite serious charges and still continues to run.

I went with 50–55 % while the community hovered around 30 %.

I could see many reasons for this to go either way: inflation, appetite for going to the movies after covid-19, etc. But more importantly, even without knowing anything about the movies in question, based on their titles it sounds like there could be some strong correlation there: if either one of the two previous movies were popular, the same crowd is likely to want to go to the next one also. So it’s not like one movie independently competes with two others, but it’s whether the fans of either of the previous two movies will watch the new one. Seems like it could happen.

I entered a rather flat distribution across countries, and it paid out. I had France at 13 % while the community had them at 6 %.

The community seemed to know a lot about this, giving very strong probabilities to some nations and weaker to others. I – ignorant as always – believed that the variance of this competition is higher than the community thinks. The variance of competitions is almost always higher than people think.

The community had assigned a probability of 7 % that it would be as extreme as it was, while I clocked in at about 5 %.

Here I miscalculated in my head. Youth employment tends to go up when people graduate from university, but for technical reasons this sometimes shows up in the statistics only when the next semester starts. At the time this question was asked, I thought the latest unemployment number referred to this student-elevated situation, which meant my forecast mainly had weight below the most recent number, as I pictured these former students would find jobs. But no, August is the student-elevated month. I was calculating with one month off.

The community put the other outcome alternative at about 30 %, whereas I had it at about 50 % throughout the question’s lifetime. I always put more weight on other outcome-type alternatives because they tend to hide many alternative resolutions that we don’t think of as possibilities up front.

And indeed, this was what happened in this question. Neither I nor the community had realised that it takes a majority of votes to win the Iranian presidential election. We worked from the assumption that it only took a plurality. So when no candidate secured a majority of votes, the election went into a run-off vote, satisfying the criteria for resolving as other outcome.

This was a fairly given winner, which the community had at about 80 %. I, lacking a lot of information, gave them the same chances as Cambridge, i.e. about 40 %. Clearly I would have done better with more information!

The community had the actual outcome at a 2 % probability, and I had it at about 9 %. I bring this question up because it highlights a typical problem with these continuous questions: the community tends to forecast a fairly narrow distribution, with tails that taper off to near-zero very quickly. A trick I’ve used for a long time is to add another component that has low weight but a huge variance. This gives the total forecast distribution fat tails, and that seems to often be the right thing to do.

I think the interpretation of why fat-tailing the forecast works is that it accounts for incorrect assumptions. In other words, if all our assumptions are correct, we might predict a fairly narrow hump on the expectation. But there’s always that 5–15 % chance that we are missing some detail, and then the answer can be just about anywhere in the broad neighbourhood of the expectation. An additional component that says “wuh-huh I don’t know it could be anything” at about a 10 % weight admits that in the forecast.