Published August 2026 · Research

The two forecasts swap places three hours before the peak

Open any daily temperature market and you are handed two forecasts for the same question. There is an hourly curve, which gives a value for every hour of the day, and there is a daily high, which is one number for the whole day. They usually disagree by a degree or so. Almost everyone picks one, trusts it, and never revisits the choice. That is a mistake, because which one is right changes during the day, and it changes at a predictable time.

The crossover

We scored both forecasts against what the settlement station finally printed, on 2,348 readings drawn from days where the maximum was watched arriving rather than already set. Everything is measured against the day's peak rather than against the clock, because the peak is what the market pays on and it lands at a different hour in every city.

In the morning the hourly curve wins comfortably. It lands in the settling bucket 43 percent of the time against 12 percent for the daily high, which is not close. Then the daily high improves all afternoon while the hourly curve does not, and somewhere between four and three hours before the peak they cross. By the last hour the daily high is landing in the right bucket 58 percent of the time and the hourly curve is back at 40 percent.

The error figures say the same thing more plainly. The daily high starts at 1.21C of mean absolute error and finishes at 0.56C, better than halving. The hourly curve starts at 0.75C and finishes at 0.81C, which is to say it starts good and stays there.

The hourly curve does not get worse. It stops getting better.

That is the part worth sitting with. A forecast that does not improve as its subject approaches is behaving strangely. Six hours out, the hourly curve is genuinely the better instrument, and it earns that: it is a shaped view of the day rather than a single number.

But it has a structural problem at the end. It reports a value for each hour, and the true daily maximum almost never lands exactly on the hour. Read the highest point on the curve and you are reading the highest hourly value, which sits below the highest actual value most of the time. The measurement shows it: the curve runs 0.42C cold in the two hours before the peak, its worst reading of the day. The daily high has no such problem, because a daily high is a statement about the day, not about a particular hour of it.

This is not hindsight

The obvious objection is that a forecast issued near the peak is not forecasting much. By mid afternoon the station has been climbing for hours, and a "forecast" that agrees with a thermometer everybody can already read has told you nothing.

We checked, because if that were true the whole finding would be worthless. For each reading we asked whether the station's running maximum had already reached the value the day settled on. Across every window in the table, the answer was 0 percent. Not a low number, zero. In none of those readings had the tape arrived at the answer yet. Whatever the late daily forecast is contributing, it is not a summary of something already visible.

What this is good for

The practical version is short. If you are reading a temperature market in the morning, weight the hourly curve. If you are reading it in the three hours before that city's peak, weight the daily high, and if the two disagree in that window, the daily high is the one that has been right.

The second use is knowing when to stop. A market where the daily high says 34 and the hourly curve says 35, at two hours before peak, is not a coin flip. On this data the daily high wins that argument about twice as often. That is not a licence to size up, it is a reason to stop treating the disagreement as noise.

The third use is the one we care about most. A model that blends both forecasts at fixed weights all day is wrong twice: too light on the hourly curve in the morning and too light on the daily high in the afternoon. Fixing that is a genuine improvement and it costs nothing but a clock.

Caveats, stated plainly

The sample is 2,348 readings but they come from 14 cities over a few weeks, not from years. The windows near the peak carry roughly 100 to 130 readings each, which is enough to see a trend this size and not enough to split it by city or season with any confidence.

Both forecasts come from the same provider. A different provider's hourly curve may be shaped differently and may not carry the same cold bias. We would expect the direction to hold, since the reason for the bias is structural rather than a quirk of one model, but we have not measured a second source.

The peak is detected from the tape, as the first hour at which the running maximum reached the value the day finished on. Days that opened with the maximum already set are excluded, because on those days "the first hour at the maximum" is just the hour our collection started.

Method

Snapshots of live temperature markets are captured every ten minutes for 14 cities, each carrying the settlement station's observations, the provider's hourly curve and daily high, and the traded ladder. A day is gradable when the tape crossed that city's local midnight, which is when these markets close. Readings after local midnight belong to the next day and are dropped before the maximum is taken, since the running maximum resets there.

"In bucket" means the forecast landed within half a degree of the settled maximum, which is the resolution these ladders trade at. Error is mean absolute error in Celsius. Bias is signed, so a negative number means the forecast read below what the day printed.

Research published by PredictionSignal. Nothing here is financial advice. Market prices move and any figure quoted was true when measured.