Some money arrives on a calendar. Retirement-plan contributions land around payday, mutual funds put new cash to work at the start of the month, institutions rebalance at the end of it, and traders position before holidays. Nobody is obliged to buy at a bad price, but if those flows are tied to dates, index returns on those dates should be higher than on ordinary days.
The idea is old. The turn-of-month effect was published in 1987, and published anomalies tend to lose about half their returns afterwards, so the expectation written down was a small effect at best.
How it was tested
Pre-registered on 2026-05-24, on SPY only, because the effect was originally measured on index returns. Five turn-of-month windows combined the last 0-3 sessions of one month with the first 3-5 of the next; a sixth variant held only the last session before a US market holiday. Buy at the open of the first day of the window, sell at the close of the last. In-sample 2014-2020, out-of-sample 2021-2026, and six variants raise the Bonferroni bar to t > 2.64.
The benchmark needed thought. The strategy trades SPY itself, so comparing it with SPY over the same window gives zero by construction. The comparison chosen was SPY over a random window of the same length: is this window better than any other window?
Where it died
One run, the same day. Raw returns were positive in all six variants, and even so none cleared the raw gate: the best t was 1.78. Against random windows of equal length, 0 of 6 passed. The best was the pre-holiday day, with 0.11% per event over 53 events. The cross-sectional phase planned for a positive result was never opened.
The record
Thesis
Turn-of-month and pre-holiday days carry systematically positive index returns because of recurring structural flows: retirement-plan contributions, mutual-fund cash deployment, institutional rebalancing and pre-holiday positioning. Nobody is legally obliged; the claim is that the flows are calendar-bound and therefore predictable.
Prior: Expected a small effect at best. The turn-of-month effect was published in 1987, and anomalies lose about half their returns after publication (McLean & Pontiff 2016).
Ariel, R. (1987). A Monthly Effect in Stock Returns. Journal of Financial Economics.
Lakonishok, J. & Smidt, S. (1988). Are Seasonal Anomalies Real? A Ninety-Year Perspective. Review of Financial Studies.
Etula, E., Rinne, K., Suominen, M. & Vaittinen, L. (2020). Dash for Cash: Monthly Liquidity Needs and the Cross-Section of Stock Returns. Journal of Finance.
Specification
signal
Five turn-of-month windows (last 0-3 sessions of the month plus first 3-5 of the next) and one pre-holiday variant (the last session before a US market holiday)
universe
SPY only — the effect was originally measured on index returns
entry
open of the first day of the window
exit
close of the last day of the window (pre-holiday — same-day close)
validation
IS 2014-2020, OOS 2021-2026, HLZ Bonferroni N=6 → t > 2.64
benchmark
Because the strategy trades SPY itself, "alpha" is the window's SPY return minus SPY's return over a random window of the same length.
Gates
Pre-registrationpass
Was the hypothesis, its parameters and its pass/fail rule written down before any data was looked at?
Raw significance (legacy)fail
Are the strategy's raw returns statistically different from zero?
OOS t > 2.64 (N=6)
variants passing0 / 6best t+1.78
Raw returns were positive in all six variants — the market drifted up.
Market-adjusted alpha (legacy)fail
Does the strategy beat SPY over the same holding windows, assuming beta = 1?
alpha t > 2.64 (N=6), alpha vs a random SPY window of equal length
variants passing0 / 6best alpha t+1.17best alpha per trade pct+0.11trades53
Best variant is pre-holiday. Alpha is measured against SPY itself (beta = 1 by construction); not re-derived with Jensen alpha.
Why it died
Turn-of-month windows on SPY do return more than zero — but no more than any other window of the same length. The best turn-of-month variant had an alpha of +0.15% per event at t +0.49; the pre-holiday day reached t +1.17, far from the 2.64 that six variants require. What looked like a calendar effect is the index's ordinary upward drift.
What it taught
For an anomaly on the index itself, the right benchmark is a random window of equal length, not zero. That comparison is what separated "positive returns during the window" (true) from "the window is special" (false).