How Deep, How Long, and Why: S&P 500 Drawdowns and Recovery Times

Research question

How often does the S&P 500 fall, how deep does it go, how long does it take to come back — and how much do the answers depend on the measurement conventions used to compute them?

Method

The core daily sample is ^GSPC from 1927-12-30 to 2026-07-02 (24,742 trading days, ~98.5 years), extended by Shiller monthly data back to 1871 for long-history context. Underwater episodes are anchored at all-time highs, with the practitioner local-peak convention computed in parallel. The study adds four elements rarely treated together:

• Counting-convention sensitivity. The same data contain 73 declines of 5%+ when anchored at all-time highs, but 339 under the local-peak convention — most “how many corrections” claims differ because of the counting rule, not the data.
• Sampling uncertainty on recovery times. Bootstrap confidence intervals around median recovery, plus recovery milestones (half / 90% / full retracement).
• A Kaplan–Meier survival view of time under water, treating the ongoing episode as censored rather than ignored.
• Four measurement conventions — nominal/real × price/total-return — applied to the same episodes.

Findings

• Frequency. 73 episodes of 5%+, 26 of 10%+, 16 of 15%+, 12 of 20%+ under ATH anchoring. The 1930s and 1940s contribute zero new-high decades; the worst daily-basis drawdown is -86.2%.
• Recovery. Median recovery is 84 days for 5%+ episodes, 433 for 10%+, and 764 for 20%+ bears — but the bootstrap 95% CI on the 20%+ median spans [666, 2310] days. Milestones are faster than full recovery suggests: the median 20%+ episode retraces half its loss 220 days from the trough and 90% within 383 days.
• Recession overlap is the watershed. Of the 26 episodes of 10%+, the 10 that overlap NBER recessions run to a median depth of -35.0% and 988 days under water; the 16 that do not stop at -14.3% and 216 days.
• Conventions change the story. With dividends reinvested, the median 20%+ underwater spell shortens from 5.4 to 2.7 years. The 1929 crash “takes 25 years to recover” in nominal price terms but 7.2 years in real total-return terms — while in real terms 2000–2013 becomes one continuous underwater spell that absorbs the 2007 crisis entirely.

What this is, and is not

It is descriptive market history intended for expectation-setting: how deep, how long, and why, with uncertainty attached. It is not a prediction model or a timing signal. The practical lesson is that no drawdown number should be quoted without its convention — the same century of data supports “73 corrections” and “339 corrections”, and both are correct answers to different questions.