Glossary

The market, finance, statistics and technology terms used on Bubble Pressure, in plain English. Where the site uses a term in a specific way, the definition says how.

The gauges

Analog episode#: A past period that resembles the present, used for comparison. The burst probability uses periods from the Nasdaq Composite since 1971 and the Nikkei 225 since 1949, after volatility scaling.
Basket#: A fixed list of listed companies tracked together. The AI basket is 17 companies with exposure to AI, chosen in 2026. Because they were chosen with hindsight, the basket is today's winners, so its past rise looks better than any basket an investor could have picked at the time.
Basket index#: The value of the basket over time, starting at 100. Each day's change is the average of the companies' price changes, weighted by their market capitalisation at the previous close, so larger companies count for more.
Bubble Pressure#: The gauge's 0–100 score: the average of six signals (valuation, overextension, capital spending intensity, complacency, rate pressure and bubble signature). Five are scored as percentiles of their own history, and the bubble signature is the share of model fits that look like a bubble. It is a measure of how stretched things look, not a probability of a crash.
Burst#: The basket closing at least 50% below its highest level so far. The burst line is the level that would take.
Burst probability#: The estimated chance of a burst within 12 or 24 months. It comes from past episodes in which a stock index had risen and then fallen about as far as the basket has now, counting how many of them went on to halve. It is shown with an 80% range.
Commentary#: A short daily summary of the gauge, written by an AI model from the figures on the page. Every figure it quotes is checked automatically against the data. It is labelled as AI-written.
Forecast log#: A daily record of each forecast, written when it is published and never rewritten. It keeps the score, the regime, the model result, the burst probability and the basket at that date, so later readers can check what was said and when.
Judgement#: A weekly AI reading of the previous seven days' news headlines. It says whether the news leans the burst risk higher, in line or lower. It never changes the published number.
Lens#: A measure shown on the page for context but not counted in Bubble Pressure, for example the price-to-earnings ratio, breadth or credit spreads.
Provisional price#: A share price taken from the last trade at 4pm New York time, before the official closing price is published. They can differ slightly. The morning update uses provisional prices, and the forecast log waits for official closes.
Regime#: A one-word summary of the basket's state. Bear market: 20% or more below its high. In correction: 10% to 20% below. Bubble pattern: at least a fifth of the model fits qualify as a bubble, or Bubble Pressure is 70 or more. Elevated: Bubble Pressure 50 to 70. Normal: anything else. The first that applies, in that order, is shown.
Signal#: One of the measures that make up Bubble Pressure. Each is scored from 0 to 100. A signal shown on the page but left out of the score, such as relative strength, is marked as such.
Watch readings#: Five conditions shown alongside the burst probability for context: the basket in a bear market, the basket 35% or more below its high, the US high-yield credit spread above 5%, the VIX 20-day average above 30, and capital spending falling at three or more of the five big spenders. They don't change the probability.

Markets

10-year Treasury yield#: The interest rate on 10-year US government bonds. Higher yields make profits expected far in the future worth less today, which weighs most on companies valued for long-term growth.
200-day moving average#: The average closing price over the last 200 trading days, roughly ten months. A price well above it suggests a strong recent run. The overextension signal measures how far the basket is above or below it.
Bear market#: A fall of 20% or more from a high.
Breadth#: How widely a rise is shared. Here, the share of basket companies trading above their own 200-day average. A rally carried by fewer companies is more fragile.
Bubble#: A period when prices rise far above what the underlying businesses seem to justify, driven largely by expectations of further rises. Bubbles are usually only certain in hindsight, after prices fall.
Concentration#: How much of a basket's value sits in its largest members. Here, the three largest companies' share of the basket's total value.
Correction#: A fall of 10% to 20% from a recent high.
Credit spread#: The extra interest companies pay to borrow compared with the US government. Investment grade covers the most creditworthy companies and high yield the least. Widening spreads mean lenders are getting nervous. The site uses ICE BofA index spreads, published through FRED.
Dot-com bubble#: The rise and collapse of internet and technology shares around 2000. The Nasdaq Composite peaked on 10 March 2000 and fell about 78% by October 2002.
Drawdown#: How far a price or index is below its highest level so far, as a percentage.
KOSPI and TAIEX#: The main stock indices of South Korea and Taiwan, home to major chip makers: Samsung and SK hynix in Korea, TSMC in Taiwan. The site shows them for context only.
Market capitalisation#: A company's share price times its number of shares: what the market values the whole company at. Often shortened to market cap.
Nasdaq Composite#: An index of almost all companies listed on the Nasdaq exchange, weighted towards technology. Its 2000 peak and crash are the standard example of a technology bubble.
Nikkei 225#: The main index of large Japanese companies. Its 1989 peak, followed by a fall of more than half, is the other burst in the burst probability's record.
Official close#: A share's closing price as set in the exchange's closing auction at the end of the trading day. It can differ slightly from the last trade.
Relative strength#: One index divided by another, showing which has done better. Here, the AI basket against the Nasdaq Composite since ChatGPT's launch.
Split-adjusted#: Prices and share counts restated so a stock split doesn't look like a crash or a jump. After a 10-for-1 split, each earlier price is divided by 10.
VIX#: Cboe's Volatility Index: the volatility of the S&P 500 that options prices imply for the next 30 days. Often called the fear gauge. Low readings mean calm markets, which the complacency signal scores as more bubble-like.

Company finances

10-K, 10-Q, 20-F and 6-K#: Reports filed with the US Securities and Exchange Commission. A 10-K is a US company's annual report and a 10-Q its quarterly report. Foreign companies file a 20-F each year and 6-Ks for news in between.
Capital spending (capex)#: Money spent on long-lasting assets such as data centres, chips, land and equipment. For AI, it measures how much is being invested in the build-out.
Capital spending intensity#: Capital spending as a share of revenue. The signal uses the five biggest spenders: Microsoft, Alphabet, Amazon, Meta and Oracle.
Hyperscaler#: A company that runs cloud computing at enormous scale, such as Amazon Web Services, Microsoft Azure and Google Cloud.
Leases signed but not started#: Leases, often for data centres, that a company has agreed to but that haven't begun. They aren't on the balance sheet yet and are disclosed in the notes to the accounts.
Operating cash flow#: The cash a company's business generates in a period, before investment. If capital spending is more than operating cash flow, the gap has to be funded by borrowing, savings or new shares.
Price-to-earnings ratio (P/E)#: Share price divided by earnings per share: how many years of current profit the price represents. It is shown as a lens.
Price-to-sales ratio (P/S)#: Market capitalisation divided by revenue over the trailing 12 months. A P/S of 10 means the market values a company at ten years of its current sales. The valuation signal uses the whole basket's P/S.
Revenue#: The money a company takes in from sales before any costs. Also called sales or turnover.
SEC EDGAR#: The SEC's public database of company filings, free to search and download.
Trailing 12 months (TTM)#: The total for the last four reported quarters. It smooths out seasonal swings.
XBRL#: A format that tags each figure in a filing, such as revenue, so computers can read it. The site takes company financials from XBRL data on EDGAR.

Models and statistics

80% range#: The range the estimate falls in with 80% probability, given the method and the data. It reflects how few past episodes there are, not every way the method could be wrong.
Backtest#: Testing a model on past data to see how it would have done. A good backtest doesn't prove a model works, because past patterns may not repeat.
Base rate#: How often something happens anyway, without any signal. A warning signal is only useful if the event follows it more often than the base rate.
Brier score#: A way to grade probability forecasts once the outcome is known: the average squared gap between the forecast and what happened, scored as 1 or 0. Lower is better. Always forecasting 50% scores 0.25.
Critical time#: In LPPLS, the date the fitted pattern points to, when a bubble is most likely to end. It is a statistical estimate with a wide range, not a crash date.
Fit and qualified fit#: Each day the model is fitted to 26 windows of past prices, from about six months to three years long. A fit qualifies as a bubble only if its parameters fall inside published limits. The bubble signature signal is the share of the 26 fits that qualify.
Hash chain#: A way of making a log tamper-evident. Each entry includes a fingerprint (a SHA-256 hash) of the previous entry, so changing any past entry breaks every fingerprint after it.
Jeffreys prior#: A standard neutral starting assumption for estimating a probability from few cases. It keeps estimates away from exactly 0% or 100% when the record is short.
LPPLS#: The log-periodic power law singularity model, developed by the physicist Didier Sornette and colleagues. It looks for prices rising faster than exponentially, with swings that speed up towards a critical time, a pattern it associates with bubbles. Its record is mixed, and the site treats it as experimental.
Out-of-sample#: Data the model was not built or tuned on. A forecast checked against later, unseen data is the strongest test. Reruns of the model on past data are not out-of-sample if the method was designed with that data in view.
Percentile score#: Where today's reading ranks in its own history. A score of 90 means the reading is higher than on 90% of earlier days. Only earlier data is used, never later.
Volatility scaling#: Adjusting past price moves so periods with different typical swing sizes can be compared fairly.
Walk-forward test#: A backtest that only lets the model see data available at each past date, stepping forward through time. It avoids using information from the future.

Data and technology

Data pipeline#: The programs that collect the data, run the models and write the results. Here it runs twice a day, at 8:45am and 4:45pm Sydney time.
FRED#: Federal Reserve Economic Data, a free database from the Federal Reserve Bank of St. Louis. The site takes interest rates, credit spreads and long index histories from it.
JSON#: A plain-text format for structured data that most programs can read. Each gauge's data can be downloaded as JSON files.

Last updated 7 November 2026. Missing a term? Suggest one.