Predictive

What the desk expects

One row per city-day still open. The bucket the model puts most weight on, what the market charges for that bucket, and the largest tradeable edge anywhere on the ladder. A row with no probability has a forecast but no priced market yet.

Depth is how many days before resolution, height is temperature, width is the resolution day. Each ribbon is one model walking from what it said a week out to what it said that morning; the green bar is what actually happened. A good model's ribbons narrow toward that bar. A biased one runs parallel to it and never arrives — and those two look identical in a table of mean error. The shaded planes are the real bucket boundaries, because a 0.6 °C miss across a line loses and a 0.9 °C miss inside one wins.

Every settled day, one dot: what was forecast a day out against what the day actually did. On the diagonal is a perfect call; above it the day came in hotter than said, below it cooler. Green is within 1 °C, which is roughly one bucket — the resolution the market actually pays at, so a dot being green matters more than it being close. Bias and error are read separately: a forecast that is 1.5 °C hot every day is a correction you can apply, and one that is 1.5 °C off in random directions is not. The numbers below split them.

The same error, split by how far ahead the call was made. A model that is sharp tomorrow and useless on Friday averages out to “fine” — this is the only place that shows. The practical use is picking the lead day at which to stop trusting it: where the line crosses one bucket wide, the forecast has stopped resolving which bucket wins.

Hit rate, per city, per lead

Bias is kept separate from error on purpose: a model 1.5 °C hot every single day is fixable, one that is 1.5 °C off in random directions is not, and pooling them into “1.5 °C error” throws away which you have. Hit rate is the only accuracy the market pays for — did the day land in the bucket the forecast pointed at.