On the S&P, Buying the 2% Dip Is Just Owning the S&P

The literal rule made money on the S&P and still failed. Buying every ES close that fell 2% or more and exiting at the next close returned $345.50 a trade across 103 trades, 57.3% of them winners — with a t-statistic of 0.90. Drawing 103 random sessions out of the same window and holding them the same way beats that more often than one time in twenty.

When we ran this test on seven years of Nasdaq futures we called dip buying real, with a shape the folk version gets backwards. Twelve and a half years of our own ES archive says the shape was Nasdaq’s and the level was the market’s.

The rule

• Daily bars from our ES tick archive, regular session, New York time.
• A session that closes 2% or more below the previous close is a signal.
• Buy that close, exit at the next close. One contract, no stop, no filter.

What “always comes back” has to beat

Holding any ES close to the next close made money over this window. That is the number the rule has to clear, and nobody who quotes the rule ever quotes it.

	Literal ≤ −2%	Every session
Trades	103	3,222
Win rate	57.3%	53.1%
Average per trade	$345.50	$60.20
Median	$308.00	$58.00
Total, one contract	$35,586.50	$193,963.50
t	0.90	1.67
p	0.37	0.095

The rule pays about six times the unconditional average per trade, and it is still the weaker of the two columns statistically, because 103 observations of a noisy thing tell you very little. Note also where the lean lives: sessions that closed up returned $30.42 held to the next close, half the all-session average. Down days do pay more. They just never pay enough to separate from their own noise.

We made that explicit with a placebo. Two thousand times, we drew 103 random sessions from the same 3,223 and measured the same one-day hold. Those draws average $63.92. Their 5th percentile is −$273.58 and their 95th percentile is $380.78 — above the $345.50 the rule actually earned. The signal sits inside its own placebo distribution.

The ladder: the shape is there, the significance is not

The Nasdaq study’s one genuinely useful result was that moderate dips revert better than violent ones, which is the opposite of what the rule’s fans believe. We rebuilt that ladder on ES: every session sorted by how far it fell, each bucket held one session forward, plotted against the flat line of what an average day pays.

Dip bucket	n	Win rate	Average	Median	t
worse than −4%	14	57.1%	−$118.79	$433.00	−0.06
−4% to −3%	25	44.0%	−$54.50	−$329.50	−0.10
−3% to −2%	64	62.5%	$603.31	$408.00	1.51
−2% to −1.5%	85	45.9%	−$58.32	−$217.00	−0.14
−1.5% to −1%	155	58.1%	$337.52	$170.50	1.84
−1% to −0.5%	324	56.5%	$217.84	$195.50	1.87
−0.5% to 0%	782	53.5%	−$17.26	$83.00	−0.26

Both deep buckets are negative. The two strongest cells by t-statistic are moderate, at t = 1.84 and t = 1.87, and neither clears two. So the direction the NQ article found is visible on the S&P and nothing more than visible. Then the −2% to −1.5% bucket ruins the gradient. It loses $58.32 a trade, and it sits directly between the two highest-paying cells in the table.

That non-monotonicity is the tell. A real effect that decays with dip size should decay in order. This one zig-zags, which is what a table of noise looks like when you cut it seven ways.

The hold sweep is a trap, and it is the important part of this article

Take the moderate bucket — every session that closed between 1% and 2% down, 240 of them — and hold it longer. The numbers get beautiful.

Hold	Win rate	Average	Every day, same hold	Excess	t	p
1 day	53.8%	$197.32	$60.20	$137.12	0.78	0.44
2 days	57.1%	$392.22	$150.63	$241.59	1.09	0.27
3 days	57.9%	$575.55	$241.53	$334.02	1.19	0.23
5 days	60.0%	$797.64	$424.54	$373.10	1.06	0.29
10 days	64.2%	$1,386.28	$881.72	$504.56	1.10	0.27

Measured against zero, the ten-day hold is the best result in this study, at t = 2.85. Measured against the same ten-day hold applied to every session in the archive, the excess carries t = 1.10. Both columns are in the table above, side by side.

The long hold is not harvesting the dip. It is harvesting the fact that the S&P went up for twelve years, and the longer you hold anything the more of that you collect. Every column in that table gets prettier as the hold extends. The excess t-statistics move around between 0.78 and 1.19, and not one of them clears its own noise. This is the most common way a dip-buying backtest lies to its author. The baseline is left out, the hold is extended until the win rate looks like a system, and index drift is booked as skill.

Twelve and a half years of both versions

Here are the two candidates as equity, one contract, costs in: the literal rule the friend describes, and the moderate bucket that the ladder liked best.

The moderate bucket finishes at $47,357.50 across its 240 trades. The literal rule finishes at $35,586.50. Both curves end higher than they start, and neither of them is a business.

The year-by-year table for the moderate bucket says why.

Year	Trades	P&L
2014	16	$1,465.50
2015	21	$3,530.50
2016	16	$3,815.50
2017	4	$494.50
2018	15	−$5,267.50
2019	11	−$824.50
2020	17	−$8,689.00
2021	16	$17,328.00
2022	42	−$32,151.50
2023	27	−$4,796.50
2024	16	−$9,084.50
2025	22	$57,726.00
2026	17	$23,811.00

Six red years out of thirteen. The worst year and the best year, both in the table, are about ninety thousand dollars apart. 2022 delivered the most signals of any year and lost the most money. Dips cluster in the years when buying them is worst, which is the mechanical reason this family disappoints. The whole twelve-and-a-half-year total is smaller than one good year in it.

What we changed

Nothing in the live book, because we never traded a fixed-percentage dip rule. What changed is the standing of the earlier article.

1. The NQ verdict gets narrowed. We said dip buying was a real edge with a shifted shape. On the S&P the same procedure produces a shape that does not hold up and a level that does not beat random days. A finding that appears on one index and not its neighbour is a finding about that index, and we should have written the NQ piece with that caveat already attached.
2. Every hold sweep in this series now ships with its baseline column. The ten-day row above would have passed an internal review a year ago.
3. Deep dips stay off the buy list. The two buckets below −3% are negative here across 39 trades, and that is consistent with what the NQ tape said about big drops carrying information.

The archive this ran on is for sale: ES ticks back to 2014 with the real aggressor side on every print, in the historical data packages.

Methodology: ES regular-session daily bars (09:30–16:00 New York) built from our own tick archive, 3 January 2014 to 10 September 2026, 3,223 sessions. Signals on the daily close, entry at that close, exit N sessions later, one contract, $4.50 commission plus two ticks slippage per round trip, $29.50 total. Baselines are the identical hold applied to all 3,222 tradeable sessions. The placebo draws 2,000 random samples of 103 sessions from the same window. t-statistics are on the trade series against zero, except the hold-sweep excess columns, which test the bucket against the all-session hold of the same length.