Stretched, but not yet breaking.
A monthly gauge of US equity bubbles since 1883. Pressure measures how stretched prices are; Trigger measures whether they have started to crack. Today the first is high and the second is quiet.
Latest reading · October 2026
Explore the history
Every month since 1883. Pick a crisis, a date range or a quadrant to see how the index behaved around it.
BFI, Pressure and Trigger, monthly
0–100, each the point-in-time percentile of its components. BFI is the geometric mean of Pressure and Trigger. Hover for a month; click one to filter to its quadrant. The strip under the lines shows each month's quadrant.
Quadrant mix in the selected period
Share of months in each quadrant and the average readings.
BFI-AI, monthly
The same recipe on an AI-era basket. Few years of history, so its percentiles move fast.
What the evidence says
Valuation is extreme
CAPE sits near the top of its range since 1881, which has historically meant low long-run returns.
Nothing is breaking yet
Trend is intact and drawdowns are small. The short-term clock is quiet.
Timing is not on offer
Four rounds of pre-registered, audited tests found the index no better than CAPE plus a simple drawdown rule.
The AI build-out
Five hyperscalers now spend most of their operating cash flow on capex, and a growing share is debt-funded. The pressure is concentrated in a handful of firms rather than spread across the economy.
Read the AI fundamentalsInside the index
Each index is two clocks built from blocks, and each block from a few measured inputs. Every number is a percentile against that input's own history, so 90 means higher than nine months in ten. Open a block to see its inputs.
BFI · US market
—BFI-AI · AI segment
—Pressure blocks: slow measures of excessTrigger blocks: fast measures of strainThe dark tick marks the line where a clock turns high: 70 for Pressure, 60 for Trigger.
Every major fall since 1929
How the index read at each market peak. Pressure often ran hot for a year or more before the top; Trigger tended to cross 60 near or after it.
| Episode | Peak | Trough | BFI at peak | Pressure | Trigger | Max BFI, 2y before | Pressure first ≥70 | Trigger first ≥60 |
|---|
Evidence
Each test was written down before it ran, run once on data it had not seen, and checked by a separate auditor. Most found nothing, and that is reported as a result.
Cycle 4 verdicts, after independent audit
Pre-registered questions, their verdicts and what the auditor changed.
| Item | Question | Verdict | Key result | Audit |
|---|
What the model implies, as ranges
5–95% range from bootstrapping the model fit; the tick is the historical base rate. The ranges straddle the base rate, so the model adds little.
Industry run-ups that later crashed
Greenwood–Shleifer–You replication on 49 US industries: percent that fell 40% within two years, overall and after a two-year run-up above 100% and 150%.
| Period | All months | Run-up >100% | Run-up >150% |
|---|
AI build-out
Capex, cash flow, debt and concentration from SEC filings. Later quarters are sealed and opened on fixed dates, so nobody can tune the measure to the answer.
Hyperscaler capex against cash flow
Microsoft, Alphabet, Amazon, Meta and Oracle, trailing 12 months.
Top-10 share of the S&P 500
Month-end, full market cap; a lower bound. Later data is held back with the sealed vault.
Sealed vault: when each quarter is opened
The AI fundamentals block was frozen before anyone saw these quarters. One boom proves nothing general, so readings stay descriptive.
| Quarter | Opens | Why this date |
|---|
Method
How the index is built, how it was tested, and where its limits are.
How it is built
Two clocks
Pressure averages slow measures of excess: valuation, credit, issuance, leverage and run-ups. Trigger averages fast measures of strain: drawdown, trend, volatility, spreads and the yield curve. BFI is the geometric mean of the two.
No look-ahead
Each input becomes a percentile against only the history known at that month, with at least ten years of data, and is lagged to its real publication date.
Four quadrants
Calm when both are low; Inflating when Pressure is high; Shock when Trigger is high; Bursting when both are high.
How it was tested
Pre-registration
Every hypothesis, sample split and pass bar was committed to version control before the test ran. Holdout data was opened once.
Independent audit
A separate reviewer re-ran every result, looked for look-ahead and leakage, and could weaken or overturn a verdict. Several were.
Known limits
Only about ten independent US crashes exist, so every probability is wide. Bubbles run on different industries each time; no single episode proves a rule.
Data
Robert Shiller (prices, earnings, CAPE), FRED and Federal Reserve Z.1, Kenneth French data library, SEC EDGAR XBRL, Jordà–Schularick–Taylor Macrohistory Database. All free and public.