Forecast Accuracy
Every forecast published on this site is written to an append-only ledger the moment it is generated, then scored against what actually happened once its 30-day horizon elapses. This page is that record.
What is being measured
We forecast a range with a probability, not a single number. So the question is not "was the price right" — it is calibration: when we say 80%, does it happen 80% of the time?
The central estimate deliberately sits close to the current price. Over a 30-day horizon a random walk is very hard to beat, and models that appear to beat it usually do so by extrapolating recent trends, which measurably increases error. We report our point error next to a random walk's so you can see we are matching it rather than claiming to beat it.
The record is still being built
41 forecasts have been logged and are waiting out their 30-day horizon. The first scores will appear here once those horizons elapse.
We are not backfilling this table with simulated results. It will stay empty until real published forecasts have actually matured — a scorecard you can only populate by waiting is the only kind worth showing.
How to read this honestly
- Small samples mean little. A coverage figure over a handful of resolved forecasts is noise; treat anything under ~30 samples per asset as indicative only.
- Overlapping horizons are correlated. Consecutive daily forecasts share most of their 30-day window, so the effective sample size is well below the raw count.
- Good calibration in a calm market says little about a violent one. Volatility estimates lag regime changes, and the bands will be too narrow until they catch up.
- None of this makes any individual forecast reliable. A well-calibrated 70% is still wrong 30% of the time, and that is the point of publishing the number rather than a headline price.
Methodology and model details are on the methodology page. Forecasts are logged daily and scored automatically; nothing on this page is entered by hand.