I think this is a good change.

One question: Have you normalized the system so that people should still expect around the same average points per question? This was one the only metrics of forecasting performance available to the public and it would be bad if it got thrown off by this change.

Community forecast on this is ridiculously high. In the GISS temperature record since 1880, the difference between the January-May average anomaly for a year and the average temperature anomaly for the entire year has mean zero, standard deviation 0.071 and excess kurtosis -0.66. The January-May average for the year 2021 was 0.782, and the annual mean temperature anomaly that has to be surpassed for 2021 to be the hottest year on record is 1.02. That means we want the difference between the two to be 0.24 or more. The base rate for this in the entire re...
[From a Georgian news outlet](https://civil.ge/archives/483425): > The Prime Minister also doubled down on his earlier statements on the refusal to join sanctions against Russia, while also stressing Georgia’s support to Ukraine in international platforms. > “At the political level, we have supported Ukraine in all formats, be it the UN, the EU, the Council of Europe, or the OSCE formats, everywhere. We were co-organizers, co-sponsors, or just supporters of the Ukrainian resolutions.” > As for sanctions, he continued, Georgia will not impose economic ...
Fitting an AR(1) to Putin's approval rating time series gives me the model \( a_{t+1} = 6.44 \% + 0.91 \times a_t \), which oscillates around a mean value of roughly 74%. Moreover, the residuals of this model have low skewness: big moves up or down seem to be about equally common. If we knew nothing about the situation in Russia, we should probably center our forecast around 74% and have a 50% confidence interval that's plus or minus 4.6% around that number or so. The inside view should obviously dominate this forecast if we knew more about it, but I'm ...

@ege_erdil It's refreshing to be proven wrong on such short order.

I think there's an important wrinkle in the resolution criteria for this question: would we say that the reign of the last Roman emperor ended in 476, or in 1453?

There's a straightforward way to calculate this kind of probability in the Black-Scholes model. People have talked before about using straddles and Black-Scholes to obtain the probabilities at a specific time horizon and then doubling it per the reflection principle, but this doubling is not exact when the Brownian motion has a drift term, which in this case it does because of the concavity of the logarithm function: in the Black-Scholes model, the underlying follows \[ \frac{dS}{S} = \mu \, dt + \sigma \, dW \] so Ito's lemma gives \[ d \log(S) = \le...
I'm surprised at the low community forecast on this question. I suspect that people are overlooking the fact that the quantity to be forecasted is GDP per capita (PPP) in *current international dollars*, not constant international dollars. This means inflation in the US is going to increase the final value along with real GDP per capita growth in Peru. Bond markets currently forecast around 2.5% annual inflation (looking at TIPS spreads) for the US over the next 10 years, and Peru's GDP per capita growth in recent years has averaged around 1.5% per year...
@(JonathanShi) Honestly I think this kind of explanation is just as bad as @alwaysrinse's rant in the top comment in this thread. The prior expected value of any kind of effect size like this is low, and the evidence is simply not compelling enough to seriously consider basing any crime rate forecast for the future on something you know about lead concentrations. If you were going to do that, I'd be happy to bet against you with a dumb time series model fit on the past crime rate data and I think I would do much better. This is true for @(j.m.)'s commen...

Similar to @SimonM's approach, but with a slightly different model and a different spot price at $48.5k. I end up with 10-90th percentiles

[58399, 58399, 58481, 63715, 70181, 78044, 89504, 105414, 138002]

The resolution criteria for this question are unclear. > A great power is a nation generally considered to have large amounts of military might and influence. While there is no established definition, for the purpose of this article, a great power is one of the top 10 nations by military spending according to the most recent report released by the Stockholm International Peace Research Institute (see latest report here). As of 2020, the great powers are therefore the United States, China, India, Russia, Saudi Arabia, France, Germany, the United Kingdom,...

In this list of 90 state lawmakers accused of sexual misconduct since the start of 2017 compiled by Associated Press, 33 of the lawmakers involved have had to leave office after the allegations against them had surfaced. That gives a naive base rate of 1 - 33/90 = 63% that Cuomo survives the scandal without being forced to leave office in some manner.

— edited by ege_erdil

I'm surprised by how little progress the Russian army has made since they first reached Kyiv. They have been bringing up troops and supplies close to the city in the past days, but even if they begin a full-scale assault now without regard for civilian casualties and with heavy artillery bombardment & air support, it's unlikely they'll be able to take the city before the end of the month. The reference class includes the two battles of Grozny in 1994 and 1999, as @pestle has pointed out. Ukrainian resistance has been stiff and Russia has been making no ...
A dumb AR(1) estimated on the log transformed past oil consumption data and then simulated forward from 2020 using Monte Carlo gives about 29% odds for this. It's not a perfect model because the residuals in the AR(1) still look autocorrelated, but correcting for the autocorrelation doesn't do much to the result, so I stopped searching for better time series models at this point. I doubt the estimate would change much even with better models. Obviously this approach is rather stupid but I don't think we have any other way of grappling with the question ...
Community forecast of 20% is surprisingly high for an event which has never happened before and without an obvious current trend in its direction that could be extrapolated to the future. If anything, there seems to be a trend in the opposite direction. The absolute maximum share of the world's population that has ever lived under one state was probably somewhere between 30% and 40%, in either a Chinese dynasty or an Indian state. Even the Mongol empire at its peak only contained around 25% of the world's population. The prediction that one state will c...
Add a "points gained over the community forecast" metric, averaged (or not, or possibly both) over all questions answered by a user. (Perhaps with standard errors as well?) The scoring system currently awards the user points for both reducing entropy relative to a uniform distribution and also being more accurate than the community. While the second term in the scoring is zero for someone who exactly mimics the community forecast for a question, the information gain term rel to the uniform distribution isn't. This means a user can reliably accumulate po...

My forecast comes from a (1, 1, 0) x (1, 1, 0, 52) seasonal ARIMA estimated on the past 5 years, where shale oil seems to have caused a shift in the past trend. The Python script along with the csv file of past weekly data to run it on can be found here. I add some additional variance to the forecast to reflect model uncertainty. The model ends up with a median of 1.305, with a 50% confidence interval (1.297, 1.313).

— edited by ege_erdil

There's a simple model proposed by Robert Lucas for the dynamics of "economic miracles" in real GDP per capita growth terms: the GDP per capita growth of a well managed country X is \[ \textrm{growth}_c = \textrm{growth}_{\textrm{leader}} \times ((\textrm{GDP per capita of leader})/(\textrm{GDP per capita of X}))^{\theta} \] where \( \theta \) is some exponent to be estimated on the data, usually about 2/3. This model describes past economic miracles remarkably well, that of Japan for example. Estimating it on the Chinese data from 1990 to 2018 gives a...

The fact that there is no restriction for the poll to be representative makes this question very likely to resolve positively. You can easily get such a result if you poll a specific subset of physicists, say, a subset working at an institute where the Everett interpretation is particularly popular.