Leaderboard
Each model was shown the same evidence and asked how tense things would be a week later. This page asks the only question that matters about those answers: were they any better than guessing?
Two reference lines decide it. A model has to beat the no-skill baseline — the error you would get from a trivial guess — and its correlation has to be large enough that it could not easily have arisen by luck, given how many days it was scored on. Being merely first in the table is not a result; the table is sorted whether or not anything in it is real.
What the two columns measure
MAE is the average size of a miss, in the units being forecast — lower is better, and 1.128 is what a no-skill guess achieves. It penalises being wrong about magnitude.
Spearman ρ asks only whether the model put the tense days above the calm ones. It ignores magnitude entirely, so it is the more generous of the two tests — a model that is uniformly too high still scores well if the ordering is right.
The critical value is computed per row from that row's own n, because models that failed on some dates have smaller samples and therefore a higher bar. Reporting one stratum-wide n would flatter exactly the models that worked least often.
No model here did better than guessing
None of these 5 models forecast better than guessing. Every correlation in the table is too small to distinguish from luck at this sample size, and the table below shows all of them so you can see how far short each one fell.
This is a finding, not a gap in the data. 5 frontier models were given identical evidence and asked what the next 14 days held, and none of them beat chance. Whichever row sits at the top is at the top of a list of results that are all indistinguishable from noise — read the badges, not the order.
Models · depth shallow · k = 14
Sorted by average error, the one column defined for every row. Sorting by correlation would put a model at the top of a column in which nothing is distinguishable from chance, and rank order is read as a verdict whether or not it is one.
| Model | Attempted | Scored | Completion | MAE vs no-skill 1.128 average size of a miss — lower is better, and 1.128 is what guessing gets you | Spearman ρ vs critical did it rank tense days above calm ones — and is that more than luck at this sample size |
|---|---|---|---|---|---|
claude-opus-5 anthropic | 55 | 55 | 100.0% | 0.8759 below the 1.128 no-skill line | 0.079chance |0.079| < 0.266 — with only 55 scored days, a correlation this size turns up by luck more than 1 time in 20, so it is not evidence of skill — how that bar is set |
claude-haiku-4-5 anthropic | 55 | 55 | 100.0% | 0.9277 below the 1.128 no-skill line | 0.174chance |0.174| < 0.266 — with only 55 scored days, a correlation this size turns up by luck more than 1 time in 20, so it is not evidence of skill — how that bar is set |
llama-3.3-70b-versatile groq | 55 | 55 | 100.0% | 0.9442 below the 1.128 no-skill line | 0.198chance |0.198| < 0.266 — with only 55 scored days, a correlation this size turns up by luck more than 1 time in 20, so it is not evidence of skill — how that bar is set |
gemini-3.1-pro-preview google | 55 | 55 | 100.0% | 0.9887 below the 1.128 no-skill line | 0.121chance |0.121| < 0.266 — with only 55 scored days, a correlation this size turns up by luck more than 1 time in 20, so it is not evidence of skill — how that bar is set |
deepseek-reasoner deepseek | 55 | 55 | 100.0% | 1.0434 below the 1.128 no-skill line | 0.091chance |0.091| < 0.266 — with only 55 scored days, a correlation this size turns up by luck more than 1 time in 20, so it is not evidence of skill — how that bar is set |
Completion rate is n_scored / n_attempted, and a model that failed stays in the denominator — the schema persists failures precisely so this column cannot be inflated by discarding them. Every attempt at this depth returned a parseable verdict, so every rate reads 100%. That is the state of the data, not a hard-coded number.
Reading the two markers together. An average error below 1.128 says the standardised verdicts sit closer to the standardised outcome than two unrelated series would; it does not say the ordering carries information. A row can slip under the error line while its correlation is indistinguishable from chance, and that combination is not evidence of skill.