Tested claim

Does the LPPLS bubble model identify tops?

A log-periodic power-law fit detects faster-than-exponential growth approaching a critical point, identifying a bubble before it bursts.

LPPLS positive-bubble confidence: qualifying windows01201320192026Qualifying windowswhen it fired
Where we are now. Today the model fits 0 of 8 windows as a positive bubble — a reading of zero means no window qualifies. Shaded bands are the periods when at least one window qualified. The 18 runs counted in the table below are defined on the fitted windows inside the model and are not reconstructable from this series.
No verdict

This rule has fired 18 times. A 40% fall followed within a year on 5 of the 18. On comparable days when it did not fire, a 40% fall followed 34.4% of the time.

PER CENT OF OCCASIONS FOLLOWED BY THE PREDICTED OUTCOMEWhen it fired28.0%When it did not34.4%18 occasions — too few for a verdict
LPPLS bubble model, scored against a transition-matched baseline
Times it fired18 separate occasions, not days
What followed28.0% a 40% fall within 365 days of that close
Happened anyway34.4% comparable days when the rule did not fire
Difference-6.5 pts near zero means the signal added nothing

Where the claim comes from

Sornette and colleagues, developed from the 1990s. We built our own implementation and publish it on this site.

How we tested it

Pre-registered Gerlach-Demos-Sornette filters, fixed before evaluation. Baseline is a seeded random-block resample over 10,000 draws.

Only days the rule could have fired are counted. Consecutive triggers within 90 days are one occasion. The baseline excludes each trigger and the 90 days after it, because the aftermath of a signal is not a fair comparison. No verdict is given below 20 occasions.

What the numbers say

561 signal days across 18 runs. A 40% fall followed 28.0% of them, against 34.4% for all eligible days. The random-block probability of seeing a result at least this good by chance is 0.67. On our specification it does not work.

What this does not mean

This is our implementation and our filter set, not a verdict on every LPPLS variant. It is the one we publish, so it is the one we test.

What would change the answer

A different filter set, a different critical-time window, or a specification pre-registered by someone else and tested the same way.

If you cannot time it, the question changes

Most of the claims in this registry cannot be shown to beat a coin flip, and several are worse than one. That is not an argument against holding bitcoin. It is an argument that the decision worth spending effort on is how much to hold, not when to buy.

We are building that second tool. Until it is here, the Where Things Stand page gives the valuation state and what has historically followed readings like today's, with the sample size attached.

Recomputed from the daily public snapshot by fetch/scorecard.py and published as scorecard.json. Method on the methods page; every claim in the scorecard.