📐 How this works
Every number on this site comes from public market data and one script you could run yourself. Here is how it is made, and where it can mislead you.
The short version
- Panic and the bottom are different things. Fear gauges (VIX, put/call, oversold breadth) say fear is extreme. They fire early and often several times on the way down.
- The bottom is confirmed by the turn, not the fear. Signals that wait for fear to drain out did better than signals that fire at the peak of fear. They cost you the first part of the rebound.
- Most lows get retested. A V-shaped bottom is the exception.
- Only extremes paid. A VIX of 40 or more beat an ordinary day; a VIX of 30 did not.
The backtest
The record behind the crash pages, the indicator pages and the track record on the dashboard comes from one script, backtest.py, run every morning after a trading day.
What counts as a correction
A fall of 10% or more in the S&P 500 from a peak, on closing prices, confirmed by a 10% rebound from the low. Until the index has risen 10% off its low, the low is not final. A bear market therefore counts as several lows: 2008–09 has five. That gives 29 corrections since 1990.
What counts as a signal
A signal “episode” starts on the first day a signal is on after 21 quiet trading days (about a month). Without that rule, a signal that stays on for weeks would be counted dozens of times. Returns are measured from the close of that first day, S&P 500 price only, without dividends, over 21, 63, 126 and 252 trading days (1, 3, 6 and 12 months).
Thresholds were fixed before looking
Every level (VIX 30 and 40, a 0.5-point credit-spread rise, 10% under the 200-day average, …) was written down from published research before the results were run, and has not been tuned since. Tuning the levels to the result is the easiest way to make a backtest look good and fail in real life.
Compared with what
Each signal is set against the average of all days in the same period. A signal that returns +10% a year when any day returned +10% has no edge, however good +10% sounds.
Did it fire in time?
For each correction we look for the first day a signal was on between the peak and 15 trading days after the low, and count how many trading days before or after the low that was. Negative means early.
The limits
- Small samples that cluster. 2008–09 alone gives several of the VIX ≥ 40 episodes, so the episodes are not independent. Thirteen episodes are not a law of nature.
- Two of the turn rules are our own constructions, standing in for breadth thrusts, which need data we only have from 2017.
- Not tested: sentiment surveys, insider buying and high-yield spreads, for lack of free history. The Baa corporate spread stands in for credit.
- US only. The S&P 500 is the yardstick; the Nasdaq Composite and Dow tabs reuse the same signal days.
- Past is not future. None of this is investment advice. It is a record of what happened, not a promise.
The dashboard’s gauges
The six panic gauges on the dashboard and on the fear calendar go back to June 2017, which covers the falls of 2018, 2020, 2022 and 2025. A gauge is “triggered” at a fixed level:
| Gauge | Triggered when |
|---|---|
| VIX | Above 30 |
| VIX term structure | Above 1.0 |
| Credit stress | 4%+ junk-bond loss vs Treasuries in a month |
| S&P 500 drawdown | −20% or worse from the 52-week high |
| Stocks above 200-day | Below 20% |
| Stocks at yearly lows | Above 15% this week |
Breadth (the share of S&P 500 members above their 200-day average, or at a 52-week low) uses today’s members, so it carries survivorship bias: companies that dropped out of the index after a crash are missing.
Headline panic and the AI
Every two minutes the site reads US market news feeds (CNBC, MarketWatch, the New York Times, the Federal Reserve and others). An AI model reads each new headline once and scores it from 0 (calm) to 100 (capitulation), with a mood and a one-line quip.
The Headline Panic Index combines the past 24 hours: 70% the average of the ten most panicked headlines, 30% the average of all of them, because fear concentrates in the loudest stories. The AI sees only the headline text, not the article. It can and does get scores wrong; every day is kept in the headline archive, so you can check.
Once an hour the AI also writes a short brief from the scariest headlines. The AI has a daily spending cap; when it is spent, scoring stops until the next day (New York time).
Data sources
- Prices (S&P 500, Nasdaq Composite, Dow, funds and the S&P 500 members): a commercial market data provider. Quotes on the dashboard are delayed 15 minutes and refreshed hourly while US markets are open.
- VIX and 3-month VIX: Cboe’s free daily files.
- Baa corporate bond spread: FRED, Federal Reserve Bank of St. Louis (series BAA10Y).
- Equity put/call ratio: Cboe’s free file, which stops in October 2019.
- Headlines: the public RSS feeds of the news sites listed above.
Times are New York time. The record updates at 07:00 after each trading day.