What "88%" actually means

One session can't prove a probability

When the low was reached at 10:44, it was tempting to say the 88% was "right." It wasn't, or at least not yet. If price had never got there, the 88% wouldn't have been "wrong" either. An 88% chance means that across many days like this one, the level gets reached about 88 times in 100 and missed about 12.

So a single outcome tells you almost nothing. A forecaster who posts only the hits is showing you the 88, not the 12.

A probability is a promise about many days, not about one.

Calibration: did 80% calls happen 80% of the time?

The way to check odds is to collect a lot of them, group them by what was said, and count what happened. If the levels called at 80–90% were reached about 85% of the time, the odds are calibrated. If they were reached 60% of the time, the number was flattering itself.

This is the odds model's history: 3,097 graded levels from 1,198 SPX sessions, December 2021 to September 2026. In every band, how often the level was reached landed within 6 points of the stated chance. Levels called at 80–90%, 85% on average, were reached 83.6% of the time. About 1 in 6 still missed, close to what the odds said.

Look at the low end too. Levels called at 0–10% were still reached 7.1% of the time. Rare isn't never.

The Brier score

Calibration needs a chart. The Brier score boils it down to one number: the average squared gap between the stated chance and what happened, counting a hit as 1 and a miss as 0.

Zero is perfect, and lower is better. A confident miss costs far more than a confident hit earns, so a forecaster can't game it by calling everything 90%.

On the 1,198 sessions, the model scored 0.169. Guessing the same average rate every day scores 0.250. On sessions it had never seen, it scored 0.173, so the edge held up.

Why a new record says nothing yet

The Monitor also keeps its own live record: every level it priced, graded after the close. Since Sept 15 that's 87 forecasts, 29 sessions of three levels each.

That's too few to judge most bands, so the Monitor holds back: it shows nothing for a band until that band has 20 graded forecasts. With two forecasts, a 50% hit rate and a 100% hit rate are the same evidence. A record that prints them anyway is flattering itself.

The model's long history and the live record are always shown separately and never added together, so 1,198 sessions of history can't stand in for a few weeks of live results.

Five questions for any odds you're shown

1. Where's the record?How many forecasts, and what happened to each one, misses included.
2. Is it broken out by band?An overall hit rate hides whether the 90% calls behave like 90%.
3. What happened at the high end?Confident calls that miss more than they should are the costly ones.
4. How big is the sample?A few weeks is a start, not proof. Look for hundreds of forecasts per band.
5. Was the number locked before the event?Odds that get revised as the day goes on can't be graded.