Does sunshine
move the market?

Hirshleifer and Shumway's "Good Day Sunshine: Stock Returns and the Weather" (Journal of Finance, 2003) tested 26 international stock exchanges against the morning cloud cover in each exchange's home city and found sunny days beat cloudy ones, on average, across the whole panel — a widely cited crack in the idea that markets run on information alone, not mood. This run isolates the single case a US reader would ask about first: does it hold for the S&P 500 and New York City, tested honestly, on the full run of data available rather than the paper's original 1982–1997 window?

Sunshine has its own enormous seasonal cycle, and so do stock returns — this desk's own runs 531 and 533 already found real May–Oct and January patterns unrelated to weather. Comparing raw sunshine to raw returns would just pit two unrelated calendar cycles against each other. So every test here ranks each day's sunshine fraction (measured sunlight ÷ total daylight that day) only against other days in the same calendar month, pooled across all 76 years — a within-month comparison that cannot be driven by the season itself. The top third of each month's sunniest days are tagged "sunny," the bottom third "cloudy," the middle third set aside for the headline test.

The full 76-year record leans the paper's direction and can't confirm it. Sunny-tercile days average +4.54 bps; cloudy-tercile days average +1.81 bps — a gap of +2.73 bps, 95% CI [-0.71, +6.17], p=0.120, Newey-West HAC agreeing (p=0.137) — contains zero. A 4,000-draw year-block bootstrap puts 95.9% of resamples on the positive side, just short of this desk's 95% bar. A continuous robustness check — every one of the 19,292 fitted days, not just the two extreme terciles, ranked by within-month sunshine percentile — agrees in size and significance: the full swing from a month's cloudiest to its sunniest day predicts +4.67 bps, p=0.060, also contains zero.

Split at 2003, the paper's own publication year, and the lean gets weaker, not stronger. Pre-2003 (8,766 tercile-days): +3.40 bps, CI [-0.47, +7.28], p=0.085 — the era the paper's own sample overlaps, still contains zero but closer to the desk's bar than the full sample. Since 2003 (4,094 tercile-days): +1.43 bps, CI [-5.49, +8.34], p=0.69 — the gap nearly disappears. A formal interaction test on whether the gap itself changed: -1.98 bps, 95% CI [-9.36, +5.41], p=0.60 — also contains zero, so the honest statement is narrow: this test cannot confirm a weather effect in either era alone, and it cannot confirm the two eras are different from each other either.

Concrete, no regression needed — and a caution about what it does and doesn't show: $1 that earns the S&P 500's log return only on sunny-tercile days and sits at 0% the rest of the time compounds, 1950–2026, to $18.53. The identical $1 restricted to cloudy-tercile days compounds to $3.19. Restricted to the 2003–2026 window alone: $3.09 (sunny) vs. $2.43 (cloudy). These numbers are real arithmetic on real data, but they are not evidence of a tradeable edge the daily regression itself failed to find — compounding a small, statistically unconfirmed daily average over thousands of trading days will always look dramatic in dollar terms even when the underlying per-day gap cannot be distinguished from noise. The dollar figures illustrate what a persistent 2.7 bp/day edge would be worth if it were real; the p-values above are the desk's actual verdict on whether it is.