We Re-Ran Our Best Strategy on the S&P — and Found the Two Years the Nasdaq Test Never Saw

We publish more kills than winners. The exception is a rule we have written about before: buy the close when the 2-period RSI drops below 10, exit at the next close. On seven years of Nasdaq futures it survived every attempt we made to break it, and it runs live as the mean-reversion sleeve of our own book.

That article ended on a sentence we now have to qualify: positive every calendar year, 2019 through 2026.

We just ran the identical rule on the S&P — twelve and a half years of our own ES tick archive, 3,224 sessions, every print carrying its real aggressor side. The first half of the result is the good half: the edge replicates on a different index almost exactly. The second half is the reason we ran it: those extra five years contain two losing years the Nasdaq test could never have seen, because our NQ archive does not go back that far.

The rule, unchanged

• 2-period RSI on daily closes, Wilder smoothing.
• Closes below 10 → buy that close.
• Exit at the next close. One contract, no stop, no filter.

Nothing was re-tuned for the S&P. Same threshold, same hold, same everything. ES costs more per round trip than NQ does in tick terms ($29.50 against $14.50, because an ES tick is $12.50), and that cost is subtracted from every trade below.

Same rule, same window, second index

	ES 2019–2026	NQ 2019–2026
Trades	192	166
Win rate	58.9%	62.0%
Average per trade	$830	$1,126
Profit factor	2.06	1.90
Sharpe	1.18	1.00
Max drawdown	$21,595	—
Days in market	9.9%	~10%

This is the result we were hoping for and rarely get. When a rule works on one market and fails on its neighbour, the honest conclusion is usually that it was fitted to the first one. Here the S&P produces the same shape — same win rate band, same profit factor band, same exposure — on a market with different constituents, different volatility and a different tick value. The edge is behavioural, and the behaviour is not Nasdaq-specific.

Now the part the Nasdaq test could not see

Our NQ tick archive starts in 2019. Our ES archive starts in 2014. Those five extra years are the closest thing to a true out-of-sample test we can run without waiting five years.

Two of them are red. 2015 lost $2,450 across 25 trades, and 2018 lost $11,032 across 32 — the worst year in the sample by a wide margin, and the year with the most signals after 2022. The 2018 damage is concentrated in the fourth quarter: the February volatility shock and the December selloff both produced clusters of oversold closes that kept going down.

So the claim has to be restated. Not “this rule has never had a losing year” — it has, twice, in a stretch of the tape that most retail backtests never load. What survives is the weaker, more useful claim: over twelve and a half years the rule made money in eleven of thirteen calendar years, and the two losses were small relative to the wins.

The twelve-year picture

	RSI-2 rule	Buy & hold
Total, one contract	$164,130	$288,950
Sharpe	0.87	0.70
Max drawdown	$23,064	$59,950
Time in market	9.6% of days	100%

Buy and hold wins on dollars. It always does on an index that went up eightfold in the sample. What the rule buys you is the ratio: more Sharpe, a drawdown less than half the size, and a flat book on nine days out of ten. Capital that is not committed to holding the S&P overnight is capital available for everything else you trade, which is the entire argument for a mean-reversion sleeve rather than a mean-reversion portfolio.

Note the shape of the blue line. Between 2014 and 2019 it goes almost nowhere: five years of work for roughly $3,500. Everything the rule earned, it earned after 2019, and a third of it came in 2020 and 2026 alone. That is the honest character of the edge — a payoff that arrives in clusters around volatility, with long flat stretches in between. A backtest that starts in 2019 makes it look like an annuity. It is not one.

What the split says

We cut the sample at 2024 and looked at both halves.

	Trades	Win rate	Average	Sharpe
2014–2023	239	59.0%	$341	0.71
2024–2026	71	60.6%	$1,164	1.32

The recent third is the strongest part of the sample, not the weakest. Whatever the rule is capturing, it has not been arbitraged away in the last two years — if anything the payoff per trade has grown with the index level, which is what you would expect from a rule that harvests a roughly fixed percentage bounce on a contract whose point value is fixed in dollars.

The variants that do not work

A rule you only ever test in its best configuration is a rule you do not understand yet.

• Every threshold from 5 to 25, every hold from one to five days, is positive on average. That is the sweep result that matters: the edge is a plateau, not a peak. If only RSI < 10 with a one-day hold had worked, we would be looking at a fitted artefact.
• The short side is dead. RSI-2 above 90, short the close: 599 trades, 43.4% wins, profit factor 0.72, −$82,308. At 95 it still loses $29,900. Overbought is not the mirror image of oversold on an index that trends up.
• The 200-day filter costs more than it saves here. On NQ, only trading dips above the 200-day average cut the worst drawdown by two thirds. On ES it cuts the trade count to 191 and the Sharpe to 0.58 without improving the drawdown at all ($21,883 against $23,064). The filter is not a free improvement; on this index it removes exactly the panic closes the rule is built to buy.

That last one matters for anyone porting the rule between markets: the filter that made the Nasdaq version safer makes the S&P version worse. Transfer the rule, not the trimmings.

What we changed in our own book

We run this logic live on NQ. Three things came out of the ES re-run that we have written into how we treat it:

1. The sleeve is sized for flat years, not for 2020. Eleven of thirteen green does not mean eleven of the next thirteen.
2. No 200-day filter when we extend the sleeve to ES. It is index-specific, and on the S&P it is a cost.
3. No short side, on either index. We had never traded it. Now we have the number that says why.

If you want to check any of this yourself, the inputs are ours and they are for sale: ES ticks back to 2014, aggressor side on every print, in the same archive this test ran on. The historical data packages carry the same files. The free GEX tool is a different question entirely, but it runs on the same feed.

Methodology: ES regular-session daily bars (09:30–16:00 New York) built from our own tick archive, 2 January 2014 to 10 September 2026, 3,224 sessions. RSI-2 with Wilder smoothing on closes. Entry at the signal close, exit at the next close, one contract, $4.50 commission plus two ticks slippage per round trip. Sharpe is computed on the trade series and annualised by realised trades per year — the same convention as the original NQ study, so the two numbers are comparable. NQ figures come from our 2019–2026 NQ daily table, which ends 27 February 2026.