Halloween Indicator (Sell in May): does it still work? Replicated through 2026
We re-ran it on S&P 500 from 2010-01-04 to 2026-11-10 (16.8 years, before costs): +7.7% a year, Sharpe 0.65, worst drawdown -33.7%. The benchmark, S&P 500 buy-and-hold, did +13.8% a year (Sharpe 0.85, worst drawdown -33.7%).
Verdict: Worse risk-adjusted return than its benchmark over this period. Historical, hypothetical results before costs, taxes and fees. Not investment advice.
This replica applies the published rule as written, on the instrument we have. It is still a backtest on ETF data, before costs.
What is lost
- One market (SPY) instead of the 37 countries studied.
- Out-of-season money is cash earning 0%, not T-bills.
Where it comes from
The origin of the 'Sell in May and go away' market proverb, with academic backing.
Rules, weights and thresholds are given as commonly cited. We have not re-checked them against the primary text.
The original rules
Stocks earn most of their returns from November to April; hold equities in the winter half-year and stay out (cash) May to October.
How we replicate it
What we changed
Applied to SPY only; the paper studied 37 countries, and out-of-season money is held as cash at 0%.
The exact rule we run
Hold 100% SPY from the close of the last trading day of October to the close of the last trading day of April; hold cash from there until the last trading day of October.
Conventions: Position decided at a close earns the next day's return; no look-ahead; before costs; cash earns 0%; starts 2010-01-04, the first day a signal exists; every market has an equal slot of equity; indicators use each market's own trading days.
Growth of 1 vs the benchmark
Benchmark: S&P 500 buy-and-hold, same period. Bought on the first day and held.
Drawdown from the previous peak
Key numbers
Sample: 2010-01-04 to 2026-11-10, 16.8 years, 4240 daily observations. Annualised from the number of observations per year of the strategy's own calendar. Before costs.
Recent and full-period returns
| Period | Replica | Benchmark |
|---|---|---|
| Last month | -0.8% | -3.7% |
| Last year | +5.2% | +11.2% |
| Last 5 years | +21.2% | +72.9% |
| Whole period (cumulative) | +251.0% | +788.9% |
Calendar-year returns
| Year | Replica | Benchmark | Replica minus benchmark |
|---|---|---|---|
| 2026* | +4.9% | +10.8% | -6.0% |
| 2025 | -4.8% | +17.7% | -22.6% |
| 2024 | +9.6% | +24.9% | -15.3% |
| 2023 | +24.6% | +26.2% | -1.6% |
| 2022 | -13.4% | -18.2% | +4.7% |
| 2021 | +16.2% | +28.7% | -12.5% |
| 2020 | +4.4% | +18.3% | -13.9% |
| 2019 | +26.0% | +31.2% | -5.2% |
| 2018 | -7.6% | -4.6% | -3.0% |
| 2017 | +11.6% | +21.7% | -10.1% |
| 2016 | +7.6% | +12.0% | -4.4% |
| 2015 | +0.5% | +1.2% | -0.8% |
| 2014 | +5.0% | +13.5% | -8.5% |
| 2013 | +19.0% | +32.3% | -13.3% |
| 2012 | +13.6% | +16.0% | -2.4% |
| 2011 | +9.7% | +1.9% | +7.8% |
| 2010 | +12.3% | +13.1% | -0.8% |
* Partial year. The first year starts on the first day of the replica; the current year is year to date.
Position as of 2026-11-10
| Holding | Weight |
|---|---|
| S&P 500 SPY | 100.0% |
| Cash (0%) | 0.0% |
Last change of position: 2026-10-30. Drawdown from its peak today: -4.6%.
What it did in 2020 and 2022
| Episode | Dates | Replica | Benchmark |
|---|---|---|---|
| COVID crash | 2020-02-19 → 2020-03-23 | -33.7% | -33.7% |
| 2022 bear market | 2022-01-03 → 2022-10-12 | -13.5% | -24.5% |
Close to close between the two dates. Two episodes only: not enough to judge a strategy.
How much to trust it
- 34 position changes over 16.8 years.
- This is one backtest on one path of history, before costs and with hindsight on the choice of strategy. It does not show the strategy will work in future.
Honest caveat
Faithful for the rule, but only one US index instead of 37 markets.
Related
Other famous strategies
Closest strategies on Quant Arena
Similar idea, tested on the site's own rule set; not the same rule.
Markets used: S&P 500
Historical, hypothetical results before costs, taxes and fees. Not investment advice.