"Sell in May and Go Away": What the Data Actually ShowsE-mini S&P 500 FuturesCME_MINI_DL:ES1!EdgeToolsEvery year around this time, the same advice circulates: sell your stocks in May, come back in November. The idea has a long history. The original saying comes from the London stock market, where traders would leave for the summer months. Academic research has studied it under the name "Halloween Indicator" since at least Bouman and Jacobsen (2002), who found statistically significant seasonal patterns in 36 of 37 countries they examined. But that was over two decades ago. Markets have changed. Algorithmic trading dominates volume, information travels in milliseconds, and the number of investors aware of this pattern has grown substantially. We tested the hypothesis from scratch, across 28 assets spanning four asset classes, with proper statistical controls. 1. The hypothesis The claim is specific and testable: returns during the six-month winter period (November through April) are systematically higher than returns during the summer period (May through October). If this is true, an investor could improve risk-adjusted returns by being invested only during winter and holding cash during summer. We define the null hypothesis as: there is no difference between winter and summer returns. The alternative: winter returns exceed summer returns by a statistically and economically significant margin. We require significance to survive Bonferroni correction for multiple testing, since we test 28 assets simultaneously. 2. The data We use adjusted close prices from Tiingo for 28 ETFs across four asset classes: 16 equity ETFs: S&P 500 (SPY, 1993-2026), Nasdaq 100 (QQQ), Russell 2000 (IWM), Dow Jones (DIA), plus country ETFs for Germany, UK, France, Switzerland, Japan, Australia, Canada, and emerging markets including Brazil, Taiwan, South Korea. 6 fixed income ETFs: US Aggregate Bond (AGG), 20-year Treasury (TLT), 7-10 year Treasury (IEF), High Yield Corporate (HYG), Investment Grade Corporate (LQD), EM Sovereign Debt (EMB). 4 commodity ETFs: Gold (GLD), Silver (SLV), Crude Oil (USO), Commodity Basket (DBC). 2 real estate ETFs: US REITs (VNQ), International REITs (VNQI). History ranges from 15 years (VNQI) to 33 years (SPY). All returns are total returns based on adjusted close prices. Fig. 1: Average monthly returns by asset across all available history. Blue shading indicates positive months, red indicates negative. Dashed lines separate winter (Nov-Apr) from summer (May-Oct) periods. 3. Monthly seasonality Figure 1 shows the full monthly return matrix. For the S&P 500, the pattern is visible but not dramatic: November (+2.5%) and April (+2.0%) are the strongest months, while September (-0.5%) is the weakest. This is consistent with the literature. The pattern is more pronounced in European equities. Germany shows July and August returns of -2.3% and -1.3% respectively, with strong November (+2.6%) and December (+2.9%). The UK shows a similar concentration of weakness in summer months. For fixed income and commodities, no clear seasonal pattern emerges from the monthly data. Fig. 2: S&P 500 average monthly returns. Blue bars are winter months, red bars are summer months. Asterisks mark months where the mean return is significantly different from zero at p < 0.05. 4. The statistical test We compute cumulative returns for each winter and summer half-year across the full history of each asset. This gives us paired seasonal observations: 34 winter periods and 34 summer periods for SPY, fewer for more recently launched ETFs. We apply Welch's t-test (unequal variances) to compare winter and summer means, following our strategy development framework. Since we test 28 assets simultaneously, the risk of finding false positives by chance alone is substantial. At an uncorrected alpha of 0.05, we would expect roughly 1.4 false positives even if the effect did not exist. We apply Bonferroni correction, setting the significance threshold at 0.05 / 28 = 0.0018. Fig. 3: Average 6-month returns, winter vs summer, for all 28 assets. Asterisks indicate uncorrected p < 0.05. Fig. 4: Statistical significance of the seasonal spread for all 28 assets. The dashed yellow line marks p = 0.05 (uncorrected), the dotted red line marks the Bonferroni-corrected threshold. Result: 4 out of 28 assets show a significant seasonal difference at the uncorrected 0.05 level: Germany (p = 0.025), Switzerland (p = 0.035), UK (p = 0.045), and France (p = 0.046). All four are European equity markets. No asset survives Bonferroni correction. Zero out of 28. 5. Effect size and confidence intervals Statistical significance alone does not tell you whether an effect is practically meaningful. Cohen's d measures the standardized difference between winter and summer returns. Fig. 5: Effect size (Cohen's d) vs p-value for all 28 assets. Points above the dashed line are significant at p < 0.05. Bubble size reflects the number of seasonal periods available. The largest effect sizes are concentrated in European equities (d = 0.5 to 0.6) and some emerging markets (South Korea, Taiwan). US equities show smaller effects (SPY: d = 0.26). Fixed income is near zero (IEF: d = 0.00). To go beyond point estimates, we compute 95% bootstrap confidence intervals using 10,000 resamples for key assets. Fig. 6: Seasonal spread with 95% bootstrap confidence intervals. Only Germany's confidence interval excludes zero. For the S&P 500, the observed spread is +2.8% with a 95% CI of . The interval includes zero, meaning the data is consistent with no seasonal effect. Germany shows +9.1% with CI , the only asset where the bootstrap interval excludes zero. 6. Return distributions Fig. 7: Box plots of 6-month return distributions for eight representative assets. Boxes show interquartile range, whiskers show the full range excluding outliers. The box plots reveal why the seasonal effect is statistically weak despite visible mean differences: the distributions overlap substantially. For SPY, both winter and summer returns range from roughly -30% to +25%. The winter median is about 3 percentage points higher, but the spread within each season is far larger than the spread between them. 7. Does it work as a strategy? We backtest a simple implementation: invest during November through April, hold 3-month T-Bills during May through October. Transaction costs of 10 basis points per trade are applied at each switch. This is a methodological improvement over many published analyses of this strategy, which assume cash earns zero during summer. In practice, an investor who exits equities would hold cash or short-term bonds, earning the prevailing risk-free rate. Fig. 8: Equity curves for the Sell in May strategy (blue) vs Buy & Hold (gray) for six representative assets. Log scale. Fig. 9: Performance summary for all 28 assets. SIM = Sell in May strategy (invested Nov-Apr, T-Bill May-Oct). For the S&P 500 over 33 years: SIM strategy: 8.0% annualized, Sharpe 0.60, max drawdown -35.7%, 49% exposure Buy & Hold: 10.9% annualized, Sharpe 0.59, max drawdown -55.2% The SIM strategy achieves a marginally higher Sharpe ratio (0.60 vs 0.59) despite lower absolute returns, because it avoids roughly half the market's volatility. It also cuts the maximum drawdown by 19.5 percentage points. On a risk-adjusted and exposure-adjusted basis, the strategy is roughly neutral for the S&P 500. The picture changes substantially by asset class and region: European equities: The strategy adds genuine value. For Germany, the SIM strategy returns 8.9% vs 6.3% buy-and-hold, with a Sharpe of 0.51 vs 0.25. The strategy outperforms in absolute terms while taking half the risk. Similar results hold for France, UK, and Switzerland. US equities: The strategy underperforms on an absolute basis but provides comparable risk-adjusted returns. For the Nasdaq 100, the strategy returns only 6.0% vs 11.0%, reflecting the strong summer rallies driven by technology stocks in recent years. Fixed income: The strategy consistently underperforms. Bond returns are more evenly distributed across seasons, and the strategy incurs unnecessary transaction costs while gaining no seasonal advantage. Commodities: Mixed. Gold shows a slight advantage for the SIM strategy (Sharpe 0.69 vs 0.61), while crude oil is negative in both approaches. 8. Temporal stability A seasonal effect that existed in one era but has disappeared is not useful for forward-looking decisions. Fig. 10: Seasonal spread (winter minus summer, annualized) for 12 assets across three uniform sub-periods. Blue indicates winter outperformance, red indicates summer outperformance. Figure 10 shows the seasonal spread across three common time windows (2006-2012, 2013-2019, 2020-2026) for 12 assets that all have data covering the entire range. The pattern is clear: most assets show positive winter spreads in the first two periods, but the picture shifts noticeably in 2020-2026. The S&P 500 spread drops from +6.0% in 2006-2012 to -3.5% in the most recent period. The Nasdaq shows an even sharper reversal, from +5.1% to -6.5%, reflecting the strong summer rallies in technology stocks since 2020. This decline is consistent with what the rolling analysis confirms. Fig. 11: Rolling 10-year seasonal spread for the S&P 500. The spread has declined from roughly 7% in the early 2000s to near zero by 2025. The rolling 10-year window for SPY shows the same trajectory as the sub-period view: a steady decline from peak spreads above 6% to values near zero by 2025. This is consistent with the efficient markets hypothesis: once a pattern becomes widely known, it gets arbitraged away. Fig. 12: Rolling 10-year seasonal spread for Germany (DAX). The spread has varied between 7% and 30% but never crossed zero, remaining consistently positive across two decades. European equities tell a different story. Germany's rolling spread (Fig. 12) peaked near 30% in 2007 and has fluctuated since, but it has never dropped below 7%. Even in the most recent window ending 2025, the spread sits around 16%. The sub-period data confirms this: Germany went from +8.6% to +10.2% in the latest period, the UK from +2.1% to +9.5%. These markets show no sign of the effect weakening, suggesting structural factors (vacation patterns, corporate reporting cycles) that may be more persistent than in the US. 9. Market regimes Fig. 13: S&P 500 seasonal spread conditional on bull/bear and high/low volatility regimes. None is statistically significant. The seasonal spread is largest during bear markets (+22.4%) and smallest in bull markets (+4.3%), but neither is statistically significant (p = 0.29 and p = 0.54). This is consistent with the hypothesis that winter outperformance partly reflects crash avoidance: three of the four largest S&P 500 drawdowns (2001, 2002, 2008) had their worst months during the summer period or early autumn. 10. Asset class decomposition Fig. 14: Average seasonal spread by asset class with standard error bars. The seasonal effect is concentrated in equities (+6.0% average spread) and commodities (+6.1%), with real estate at +5.7%. Fixed income shows essentially no seasonal pattern (+0.6%). This is what we would expect if the effect is driven by equity risk premia and investor behavior rather than a fundamental economic mechanism. Bond cash flows are contractual and do not respond to seasonal sentiment shifts. Within equities, the effect is strongest in European and emerging markets, weaker in the US, and essentially absent for the Nasdaq 100. The Nasdaq's non-result likely reflects the dominance of technology stocks, which have delivered strong returns throughout the calendar year over the past two decades. 11. Limitations ETF histories are short. Most ETFs in this study launched between 1996 and 2010. The analysis covers at most 33 years and as few as 15. Calendar effects observed over short samples can be statistical artifacts. Survivorship in ETF selection. We test ETFs that exist today. ETFs that failed may have had different seasonal patterns. This is a mild form of selection bias. No look-ahead in data construction, but in ETF choice. The ETF universe was chosen based on current availability and liquidity, not based on pre-2026 information about which ETFs would show seasonal effects. The broad coverage across asset classes mitigates this concern. Transaction costs are conservatively estimated at 10 basis points. For large institutional investors, costs would be lower. For retail investors trading options or using leveraged products, costs could be higher. The T-Bill rate during summer months is approximated using the 3-month T-Bill from FRED (series DTB3). Actual money market returns may differ slightly. 12. Summary Across 28 assets, four asset classes, and up to 33 years of data: The seasonal effect exists in the raw data for most equity markets. Winter returns are on average 6 percentage points higher than summer returns for equities. The effect does not survive multiple testing correction. Zero of 28 assets are significant after Bonferroni correction. Four European equity markets are significant at the uncorrected 0.05 level. The effect has been declining over time in the US. The rolling 10-year seasonal spread for the S&P 500 has dropped from 7% to near zero. As a strategy, it is roughly risk-neutral for US equities when cash earns the risk-free rate. The improved Sharpe ratio from reduced volatility is offset by foregone returns. European equities are the one area where the strategy has generated genuine alpha, both in absolute terms and on a risk-adjusted basis, and this has been relatively stable over time. Fixed income shows no seasonal effect. Testing it there was a waste of statistical power. For most investors, the evidence does not support leaving the US equity market in May. The effect is real but too small and too variable to generate reliable outperformance after accounting for the risk-free rate and multiple testing. European equity investors have a somewhat stronger case, but even there, the effect should be viewed as one input among many rather than a standalone strategy. References Andrade, S.C., Chhaochharia, V. and Fuerst, M.E. (2013) '"Sell in May and Go Away" Just Won't Go Away', Financial Analysts Journal Bouman, S. and Jacobsen, B. (2002) 'The Halloween Indicator, "Sell in May and Go Away": Another Puzzle', American Economic Review Jacobsen, B. and Zhang, C.Y. (2013) 'The Halloween Indicator, "Sell in May and Go Away": Everywhere and All the Time', Available at SSRN: 2154873. Kamstra, M.J., Kramer, L.A. and Levi, M.D. (2003) 'Winter Blues: A SAD Stock Market Cycle', American Economic Review Maberly, E.D. and Pierce, R.M. (2004) 'Stock Market Efficiency Withstands another Challenge: Solving the "Sell in May/Buy after Halloween" Puzzle', Econ Journal Watch Lucey, B.M. and Zhao, S. (2008) 'Halloween or January? Yet another puzzle', International Review of Financial Analysis