Skip to Forecasting Lab content
FORECASTING LAB VOL. 01

A LONG-TERM EXPERIMENT IN BETTER JUDGMENT

Forecasting Lab

The Future,
Quantified.

What do I believe will happen? How confident am I?
And how often am I right?

A public record of my predictions, uncertainties, and changing beliefs.

公开记录我对未来的判断，并让时间检验这些判断。

Explore the archive
ROOM FOR UNCERTAINTYP ∈ [0, 1]
The shape of uncertaintyAn illustrative probability density, not a personal forecast or measured performance. 0%50%100%
Certainty is rare.
Honest probabilities matter.CONCEPTUAL ILLUSTRATION
NO FORECAST DATA

An open record.

Every forecast counts. Especially the wrong ones.

Total predictions3Published, never drafts
Open predictions3Still taking shape
Resolved predictions0Outcomes on the record
Average Brier score ⓘN/AInitial forecast · lower is better
CalibrationNot enough dataAt least 20 resolved forecasts

01 / THE ARCHIVE

Questions for the future.

A growing, public record

3 published predictions

OUTCOMES, WITHOUT EDITING THE PAST

Resolved & reflected.

No resolved predictions yet. Each outcome will include evidence, its original probability, its last legal update, and a candid reflection.

A NOTE FROM YOUR PAST SELF

On This Day

What did I predict years ago?

Historical anniversaries appear here as the archive grows.

02 / KEEPING SCORE

Confidence meets reality.

0 scored forecasts

When I say “70%,” does it happen about seven times out of ten? Calibration asks whether confidence is earned.

Not enough resolved predictions for meaningful calibration analysis. At least 20 scored predictions are required; sparse buckets are labeled individually.

Reliability diagram

Initial probabilities
Ideal calibration Observed frequency · 20% bins

Performance over time

Cumulative average Brier · lower is better

The first resolution starts the curve.

Performance by category

No scored forecasts yet. Category comparisons will appear as outcomes arrive.

03 / TRY THE MATH

How much does
confidence cost?

A good forecast leaves room to be wrong. Move the slider, choose an outcome, and watch the Brier loss change.

Interactive example · outside the official record

Brier = (p − y)²

Impossible · 0%Certain · 100%
What actually happened?
Your Brier loss0.2500A 50% forecast has the same loss for either outcome.

0 is a perfect forecast. 1 is confidently wrong. This simulation never changes the archive.

04 / THE GROUND RULES

A record you can question.

METHOD / V1.0
01

Define before predicting.

Every published question has a deadline, eligible sources, and explicit Yes / No criteria. Ambiguity is addressed before scoring.

02

Keep the first judgment.

Official Brier scores use the initial probability. Belief updates are appended with reasons and evidence; they never rewrite the starting point.

03

Let the evidence settle it.

Only formally resolved binary outcomes count. Drafts, open questions, pending reviews, and annulled records never enter the score.

Scoring, sample size & provenance

For each eligible forecast, loss = (initial probability − outcome)², where Yes = 1 and No = 0. The official score is the unweighted mean of these losses. It is N/A until at least one eligible resolution exists.

Calibration uses five fixed bins: [0%, 20%), [20%, 40%), [40%, 60%), [60%, 80%), and [80%, 100%]. Analysis needs 20 resolved predictions; a bin needs five observations for interpretation. Smaller samples remain visible as descriptive data.

An event is never marked No simply because it has not happened yet. No outcomes require the resolution deadline to have passed. All resolutions need evidence and reflection.

Each published forecast preserves an original content snapshot, with subsequent updates recorded separately. This version provides no independent timestamp proof. Git dates and a standalone SHA-256 digest are not independent, tamper-proof timestamp evidence.

Visitor probabilities remain on the current page. No login, server submission, or community statistics are involved.