Weekly Mean Reversion Strategy: 22,680 Backtests on Index Futures — and the Filter I Published Was Overfit

TL;DR — the answer box

• The plain rule works, modestly. Buy the week after a down week, sell that week's close: profitable on all six markets, 54.4% to 58.2% win rates, and 3,524 signals across 19 to 33 years of data.
• The filter I published was overfitting. Volume Oscillator (95,100) > 0 ranked 1st of 139 period pairs on ES and 137th of 139 on IWM. On 5 of 6 markets it scored worse than using no filter at all, and on DIA 0 of 139 pairs beat no filter.
• It beats holding on risk, never on money. Smaller worst drawdown on 6 of 6 markets, better return-per-drawdown on 4 of 6 — and less profit than buy-and-hold on all 6. On SPY: a 19.58% worst fall against 55.91%, for $93,273 against $551,738.
• The bounce is fast and small. Price traded back up through the down week's own close inside one week in 96.5% of ES cases. The average gain that week was 0.220%.
• A higher win rate is not a better strategy. Holding 21 weeks instead of 1 lifted the ES win rate from 58.1% to 75.6% — and moved return-per-drawdown from 4.33 to 7.05 on ES while it fell from 5.53 to 4.12 on SPY.

How we tested

Weekly bars, built from daily data by taking each week's first open, highest high, lowest low, last close and total volume. Six markets: E-mini S&P 500 (ES) and E-mini Nasdaq 100 (NQ) futures from 2007 to 2026 — 1,017 weekly bars each — plus SPY back to 1993 (1,741 bars), QQQ, DIA and IWM. The Dow appears as the DIA ETF rather than YM futures, because this dataset holds no YM series; that substitution is stated here rather than buried.

The event is one comparison: this week's close below last week's. That happens in 42.3% to 44.9% of all weeks depending on the market — 3,524 down weeks in total.

Then 22,680 backtested variants, 3,780 per market. What varies: direction (long and short), the hold (1 to 21 weeks of exposure), how many consecutive down weeks you insist on (1, 2 or 3), how big the drop was in units of the 20-week average true range — ATR, a standard gauge of how far a market has been moving lately — the two regime filters the engine always applies, and the 2025 volume filter on or off.

Three assumptions, stated every time because they matter:

• $35,000 starting capital, no compounding. Futures trade one contract; ETFs buy as many shares as the account holds.
• Frictionless. No commission, no slippage. On ES a round turn runs about $17 all-in against an average trade of $397, so it is a real deduction and a second-order one.
• No look-ahead. You cannot know a week closed lower until it has closed. Every fill here is the next week's open.

That last point is a small correction to my own 2025 article, which said "go long at the close." You can do that with a market-on-close order. Most people reading it will place the trade over the weekend, which means Monday's open — so that is what I tested.

That is deliberately not the best trade in the study. It is the middle one. Half of the 434 ES trades did better and half did worse, and a picture of the best one would tell you nothing except that good weeks exist.

Does buying a down week on index futures actually work?

Yes. Small, consistent, and present on every market tested.

Ret/DD is net profit divided by the largest dollar drawdown — how much you made for the worst hole you sat in. Worst drawdown is the biggest percentage fall in the account.

The mechanism is visible underneath the P&L. Across all six markets, 99.3% to 99.7% of down weeks eventually traded back up through their own close, and 96.1% to 97.2% did it inside the very next week. The snap-back is real and it is quick.

Is the bounce a known effect, or just ours?

Known, and other people measure it a different way and land in the same place. TradingStats tracked 565 weeks of Nasdaq futures from 2015 to 2025 and found price returned to the weekly opening price in 69.7% of them. Different instrument, different reference level, different decade weighting — same phenomenon. What their study stops short of is what an account trading it would have done, which is where this one starts.

How big is it, honestly?

It is thin. The average one-week gain after a down week is 0.220% on ES and 0.351% on SPY. Stretch to six weeks and you get 1.011% and 1.337%. That is an edge. It is not a windfall, and any article that makes it sound like one — including mine — is selling.

One more thing worth saying plainly, because it looks like evidence and is not. The short side of this grid loses exactly what the long side makes, to the dollar, on every market. That is arithmetic, not a discovery: the study is frictionless, both sides see the same events, and position size does not depend on direction, so the short column can only ever be the long column with a minus sign. The control that actually tests something is the seeded random one, and I come to it below.

Does the volume filter really double your risk-adjusted return?

No. And this is the part of the 2025 article I am retracting.

Here is what I published then, on ES: adding Volume Oscillator (95,100) > 0 lifted the return-to-drawdown from 2.7 to 7.1 and the win rate from 59.0% to 63.3%. A volume oscillator is just the gap between a fast and a slow average of traded volume, so "above zero" means recent volume is running hotter than its own longer-run level.

What happens when you retest it as published

Retested in this study, at that exact setting, the filter looks even better on ES than it did in 2025 — 434 trades become 185, and Ret/DD goes from 4.33 to 9.57. If I stopped there I would be repeating the mistake with better data.

So I did not stop there. I held the rule completely still — long, one down week, every drop size, no regime filters, one week held — and swept only the oscillator's two period settings across 139 combinations from 5 to 150 weeks, on all six markets. 95 and 100 became one cell among many, scored by the same code as the rest.

What happens when you sweep it

Read the last column. The setting I published is the best available on the one market I published it for, and ordinary-to-terrible everywhere else. Read the third column too, because it is worse: on 6 of 6 markets the median filtered pair scores below simply not filtering, and on 5 of 6 markets the published pair itself scores below not filtering. On the Dow, not one of the 139 settings beat leaving the filter off.

Why five weeks is the tell

A filter fitted to one market is a suit cut for one man. On him it is superb. On anybody else it is a costume — and the giveaway is always the same: the tailoring is far too specific to be about clothes in general. Ninety-five weeks against one hundred weeks is a five-week difference measured across two years of volume. There is no market mechanism that turns on at 95 and off at 90. What there is, is a search that ran over a lot of settings and reported the winner.

That failure has a name and a literature. Bailey, Borwein, López de Prado and Zhu formalised it as the probability of backtest overfitting — the odds that the best-looking configuration in a search is best because of the search, not because of the market. It is the reason this newsletter is called Overfit. It is also, evidently, not a thing you become immune to: I lost roughly $270,000 early in my career to overfitting and leverage, and I still shipped a 95/100 volume filter in May 2025.

Does the strategy actually beat buying and holding?

On risk, yes. On money, never. And the 2025 article claimed "beat the market" while showing no market at all.

CAGR/DD is the plain calendar return per year divided by the worst percentage drawdown. It is the only ratio in this study that may be set beside buy-and-hold, because the ratio the 2025 article used annualises by trade frequency — which puts a 434-trade strategy and a one-trade hold on completely different scales and can reverse the answer.

What the comparison actually says

The honest summary: the strategy is in the market 43% of the time, takes a smaller worst hit on all six markets, wins on risk-adjusted return on four of them, and makes less money than doing nothing on every single one. On SPY it earned $93,273 against $551,738 — under a fifth of the money, for a third of the drawdown.

And it clears the luck bar, which is the control that counts here. A seeded random entry, matched to fire as often as this study's own median variant and hold as long, averaged over ten seeds, produced $88,465 on ES with a spread of ±$56,996, and $9,037 on SPY with a spread of ±$15,159. The rule's $172,488 and $93,273 sit outside those ranges. Not miles outside — outside.

Read the ETF rows, not the futures rows

One warning about the futures rows, because the drawdown numbers there are mostly about leverage rather than about the strategy. One ES contract is $50 per index point, per the CME contract specification, which at recent prices is about $370,612 of index exposure — 10.6 times a $35,000 account. NQ is 16.8 times. That is why buy-and-hold on one ES contract shows a 112.82% worst drawdown: above 100% is not a rounding artefact, it means the account was gone, in March 2009, and then some. If you want to read the risk of this idea rather than the risk of that leverage, read the ETF rows.

How long should you hold, and does waiting for a deeper drop help?

Longer holds win more often and pay no better. Deeper drops help on one market and hurt on another.

Hold for 21 weeks and the ES win rate reaches 75.6%. It is also the row with 45 trades, and its risk-adjusted return goes up on ES and down on SPY. This is exactly why a win rate on its own is a marketing number: the two markets disagree about whether the longer hold was worth anything, and only the risk-adjusted column tells you that.

Does waiting for two or three down weeks help?

It does not. Waiting costs more than it buys:

Every market gets worse as you demand a deeper hole. Drop size is the more interesting cut, and it splits: on SPY the middle band pays best (216 trades at 6.23 Ret/DD for a 0.5–1.0 ATR drop, against 2.72 for the smallest drops), while on ES the biggest drops look best but rest on 42 and 25 trades — both below the 50-trade reliability floor, both flagged, and neither strong enough to build on.

The direction regime filter is the one that earns its place. On SPY, taking the signal only when the 100-week average was rising cut trades from 761 to 578 and lifted Ret/DD from 5.53 to 9.93; taking it only when that average was falling left 136 trades at 0.37. Buy the dip in an uptrend. Do not buy the dip in a downtrend. The data is not subtle about this, and the same shape shows up in the buy-the-dip signal rankings I ran across 40 different entry rules.

The verdict — and the honest limits

The 2025 article got the idea right and the evidence wrong.

Where it was right: a once-a-week mean reversion rule on US equity indexes is real. It survives on six markets over 19 to 33 years, it clears a frequency-matched random control, it takes about 22 decisions a year, and it does the thing a defensive strategy is supposed to do — a smaller worst drawdown than holding, on every market tested. The stable region backs this rather than one lucky cell: of the reliable long variants in the grid, 96.5% on ES and 97.4% on SPY finished profitable, with median return-to-drawdown of 2.29 and 3.00.

Where it was wrong: the headline finding. The filter that supposedly doubled the risk-adjusted return was the best of 139 settings on the one market it was reported on, and it does not transfer. It also claimed "beat the market" without ever putting the market on the page, and it quoted "$182,875 net profit" without a capital base, a sizing rule or a drawdown beside it — which makes the number decoration rather than information.

The limits, because a study that hides them is not worth trusting:

• Frictionless. No commissions, no slippage. About $17 a round turn on ES against a $397 average trade — call it 4% of the average trade, gone.
• Leverage, not strategy, drives the futures drawdowns. One ES contract is 10.6 times a $35,000 account. Nobody should trade it that way; the ETF rows are the readable risk picture.
• The grid is mostly thin. 19,462 of 22,680 variants (85.8%) fall below the 50-trade reliability floor. They are flagged and kept, not deleted, and every headline number above comes from a variant above it — but a deep grid produces a lot of cells you must not read.
• Two markets say the opposite. On QQQ and IWM, buy-and-hold beat the strategy on risk-adjusted return. Four out of six is a majority, not a law.
• US equity indexes only. Nothing here licenses a claim about gold, oil, currencies or single stocks. The upward drift that makes buying weakness work is a property of these indexes.
• A backtest is not a live edge. It is the best available evidence about the past and no promise at all about the next 19 years.

What this means for you

1. Trade the plain rule or none of it. A weekly close below the previous weekly close, long only, out at the following week's close. No filter. The filter version costs you half your trades and, on five markets out of six, pays you less per unit of risk for the privilege.
2. Add the trend regime, not the volume one. Only taking the signal when the 100-week average is rising moved SPY from 5.53 to 9.93 return-to-drawdown across 578 trades. That is a real, broad improvement, not a lucky cell.
3. Size it as risk management, not as a return engine. It earned less than buy-and-hold on all six markets while taking a smaller worst hit on all six. If you want the return of holding, hold. If you want to be in the market 43% of the time with a shallower hole, this is what that costs.
4. Never trade one contract on $35,000. That is 10.6 times leverage. Buy-and-hold at that size wiped the account out in 2009 in this very test.
5. Before you trust any filter, sweep its settings. One number is not a result. If a setting only works at one value on one market, you have found a search artefact. That test took me two minutes and eleven years too long — every entry signal I rank in the 36 mean-reversion setups library now gets it before it gets published.

Two down weeks in a row is not a better signal. One volume oscillator is not a filter. And the thing that separates a strategy you can trade from a chart you can admire is whether the good result has neighbours.

If you want the studies as I finish them — including the ones where I have to correct myself — that is what the newsletter is for: StatOasis.com/Overfit

Methodology and honesty footer: 22,680 backtested variants plus 834 filter-sweep backtests, across ES and NQ futures and the SPY, QQQ, DIA and IWM ETFs, on weekly bars from 1993–2026. Flat-only, next-open fills, $35,000 fixed capital, no compounding, frictionless. Buy-and-hold and a seeded frequency-matched random control computed on the same bars. Every figure in this article is generated from the study's own results table. This is research, not investment advice.

The version of this article published on 2025-05-02 reported a volume-oscillator filter as a headline finding. That finding is withdrawn above; the original page is preserved in the archive record for this study.