Who predicted past bubbles correctly? A track record to weigh AI bubble forecasts

Why does a track record matter for AI bubble forecasts?

A forecast has no value on its own until you know how often its author is right. A record turns trust into a number. Two forecasters can give the same warning today, and the one with 5 documented hits and 1 miss deserves more weight than the one whose last hit dates back many years and whose misses went unrecorded.

The AI bubble debate raises the stakes, because a single warning can move budgets. If you know how to read a record, you can decide in minutes whether a warning is worth a second look. The method below needs a spreadsheet and the patience to find dated sources. Each entry you add later makes the next judgment faster and fairer, and the ledger can be shared with colleagues.

What counts as predicting a past bubble correctly?

Set the rules before you read the calls. A hit needs a claim made before the fall, in writing, with a date. It names an asset or a market, a direction, a rough level or size of the fall, and a window of time. A claim that a market is expensive, with no window, cannot be graded.

Grade the window with care. A call that arrives 10 years early and a call that arrives 3 months early are different results, and a fair ledger shows how much an investor could have lost by acting on each. Partial credit works well: right direction with wrong timing earns half a point. Wrong direction earns none, and a call made after the fall is excluded from the ledger.

How does survivorship bias distort who predicted past bubbles?

Stories about past bubbles feature the people who called them. The people who called many bubbles that never came are absent from the stories. This is survivorship: the winners remain visible and the rest disappear, so the visible sample overstates the skill of the group. Every list of famous forecasters carries this distortion.

Correct for it by starting from the events and not from the people. List the bubbles first, then search for every public warning in the years before each one, from all authors, hit or miss. The resulting hit rate describes the crowd, and individual hit rates now have a baseline to beat. The comparison shows whether a name stands out or simply belongs to a large group of lucky guesses.

Why are repeated bubble warnings weak evidence of skill?

A forecaster who warns of a bubble every year will eventually be right. A stopped clock is right twice a day. The question to ask is how many warnings preceded the one that came true, and what a reader who acted on each warning would have lost while waiting.

Count the cost of being early. If a warning arrives 4 years before the fall, a reader who sold at once missed 4 years of gains. The ledger should record that cost next to the hit, because it measures usefulness for a decision, which is the only use that matters to you.

How do you build a ledger of past bubble predictions?

Use one row per call and 7 columns: author, date, market, claim, window, outcome, and source link. Fill the first six from the original text, never from a later summary. Prefer primary documents such as the author’s own letter, report, or interview, because summaries drift toward the version that sounds wise.

Add a column for the stated confidence. A call made with strong confidence and a miss weighs more than a call made with open doubt. Add a final column for the reason the author gave. Reasons that rested on a measurable condition teach you more than reasons that rested on a feeling, since conditions can be checked again in the next cycle.

How do you read a ledger of bubble forecasts for AI?

Read three numbers: hits, misses, and early or late timing. Then read the sample size. Bubbles are rare, so even an excellent forecaster may own only a handful of graded calls. With so few cases, a hit rate is a rough guide and not a measurement, and the right response is modest confidence and a readiness to update as new calls are graded.

Then compare the reasons with the AI case. If a past call rested on easy credit and rising leverage, check whether those conditions exist in AI today. A good record in a different setting transfers little. The EQOS Asia AI study (April 2026) gives conditions you can check: valuations up ×4 in 18 months and 95% of projects failing in production.

How can you apply a track record to today’s AI bubble forecasts?

Take the three loudest AI bubble forecasts of the quarter and enter them in the ledger today, with the date and the stated window. Review them when the window closes. After a year you hold your own record, and it will tell you which sources deserve your attention. This page is general information, not investment advice.

A method that anyone can recompute needs no reputation. The EQOS Asia method page shows how the φ(S) score is built, and the Λ simulator lets you test your own case. For a decision in Singapore, ASEAN, or Hong Kong, a 30-minute call with EQOS Asia is the fastest way to apply it.

Frequently asked questions

About the author

Stéphane Jambu

President, EQOS Forecast Asia

Stéphane Jambu leads EQOS Forecast Asia and works as a consultant in e-commerce, search visibility, and AI. LinkedIn profile