Forecasting HN · Learning tool
Brier & Calibration Sandbox
Interactive scoring sandbox · Not HN's track record
Try the scoring rules HN uses on numbers you make up. Nothing you enter here is recorded, sent anywhere, or added to HN's ledger. Every calculation runs in your browser.
HN ledger snapshot, November 6, 2026: Open: 3 · Resolved: 0/15 — phase collecting. No headline Brier score is published until 15 questions resolve.
One forecast, one outcome
Brier score = (p − o)², where p is the probability of YES and o is 1 if YES happened, 0 if not. Lower is better: 0 is perfect, 1 is the worst possible.
| Forecast | Brier | Note |
|---|---|---|
| You | — | — |
| Always 50% | 0.2500 | Constant baseline; scores 0.25 whatever happens |
| Always the base rate q | — | — |
Expected climatology Brier q(1 − q) is the long-run average if events really happen at rate q and you always say q. It is a benchmark under that assumption, not a law.
Many forecasts: Murphy decomposition
Enter one forecast per line as probability, outcome (e.g. 70, 1 or 0.7, yes). Forecasts are grouped by their exact value, so the identity Reliability − Resolution + Uncertainty = Brier holds exactly.
Probabilities as 0–1 or as percentages (any value above 1, or ending in %, is read as a percent, so 1 means 100%). Outcomes as 1/0 or yes/no. Up to 2,000 lines.
| Forecasts (N) | — |
|---|---|
| Brier score | — |
| Reliability (lower is better) | — |
| Resolution (higher is better) | — |
| Uncertainty ō(1 − ō) | — |
| REL − RES + UNC | — |
| Skill vs. always 50% | — |
| Skill vs. climatology | — |
| Forecast | n | Observed YES |
|---|
Updating in log-odds
New evidence multiplies the odds by its likelihood ratio (LR): log-odds(after) = log-odds(before) + ln(LR). An LR above 1 favours YES; below 1 favours NO.
| Prior log-odds | — |
|---|---|
| + ln(LR) | — |
| Posterior log-odds | — |
| Posterior probability | — |
| As HN would publish it | — |
HN house rule: whole percentages from 1% to 99%, never exactly 50%. A 0% or 100% prior cannot be moved by any finite evidence, which is one reason HN never publishes them.
Why HN waits: the 15 / 50 / 150 gates
A Brier score from a handful of questions is mostly luck. HN fixes in advance how many resolved questions it needs before it shows each kind of number.
- 15 resolved — headline Brier score is shown. Below this the ledger is in phase
collectingand the headline is suppressed. - 50 resolved — calibration chart loses its early label.
- 150 resolved — phase
measured: the record is described as a measured track record.
These are presentation gates, not guarantees of accuracy. They stop an early lucky (or unlucky) streak from being read as skill.