Larry Connors Double 7 Strategy: The Exact Rules, Tested 4,320 Ways on 33 Years of S&P 500 Data

TL;DR: the answer box

• The rules, in three lines. Long only. The close must be above its 200-day simple moving average. Buy when the close is the lowest close of the last 7 days. Sell when the close is the highest close of the last 7 days.
• It works. On SPY from 1993 to 2026: 322 trades, 74.5% win rate, $75,616.50 on a $35,000 account, frictionless, with an average hold of 7.55 sessions and only 29.0% of the time in the market.
• It is not a buy-and-hold beater. The index compounded at 8.82% a year against the strategy's 3.51%. The case for the strategy is the other column: a worst fall of 8.5% against the index's 56.5%, and a CAR/MaxDD, the metric engine's return per unit of drawdown, of 0.76 against 0.22.
• Seven is not magic. Sweeping the lookback 2 to 21 days, the published 7 ranks third of eight. The 5-day version scores 0.99 against 7's 0.76. Four, five and seven all clear 0.70. That is a plateau, which is the reassuring answer, not the disappointing one.
• The clever exit is nearly ornamental. On identical entries, a plain 10-bar hold scores 0.74 with a 7.5% drawdown against the published 7-day-high exit's 0.76 and 8.5%.
• The 200-day filter costs money and buys sleep. Removing it ADDS $9,143.90 of profit and more than doubles the worst drawdown, from 8.5% to 19.6%.
• The edge decayed after publication. Average trade 0.81% across the 141 trades before the 2008 book, 0.57% across the 181 since; profit factor 2.82 falling to 2.02.
• It clears a random baseline by a distance. A random control matched to the grid's median variant made $9,146.50 at a 56.89% win rate and scored 0.040. The rule scored 0.76.

How we tested

If you have ever bought a dip in a market that was clearly still going up, and then sat there wondering whether you had a strategy or a habit, this one is for you.

The case for this rule is usually made from one equity curve. This is what happens when it is made from a grid instead, and two of the conclusions that survive are not the ones you would expect.

The setup, in plain English. Larry Connors and Cesar Alvarez published the Double Seven in Short Term Trading Strategies That Work (TradingMarkets, 2008). It is a mean-reversion rule, which means it bets that a market that has dropped a little will bounce rather than keep falling. Three conditions. First, only trade when the S&P 500 is above its 200-day simple moving average, which is the average closing price of the last 200 sessions and the standard rough line between an uptrend and a downtrend. Second, buy when today's close is the lowest close of the last seven days. Third, sell when today's close is the highest close of the last seven days. That is the whole thing. Two sevens, hence the name.

One detail decides whether a backtest of it is honest. Connors buys the close that makes the seven-day low. You cannot know at 15:59 that the closing print will be a seven-day low, so every trade here fills at the next morning's open instead. That is one bar more conservative than the book, no close-fill version was run here for comparison, and it is the difference between a number you can trade and a number you can only publish.

We tested it on SPY daily data from 1993-02-02 to 2026-06-12, which is 8,398 bars and 33.4 years, and again on E-mini S&P 500 futures from 2007-01-03 to 2026-08-17. The engine found 851 SPY signals and ran 4,320 backtest variants per instrument: two directions, eight lookbacks (2/3/4/5/7/10/14/21 days), the 200-day filter on and off, fifteen different exits, and a 3×3 grid of volatility and trend regime filters. Account: $35,000, whole-share sizing, no compounding, one position at a time, frictionless unless a section says otherwise. Every number below comes from the study's locked facts file, computed straight from the result CSVs.

Does the Connors Double 7 strategy still work on the S&P 500?

Yes. Here is the rule exactly as published, with nothing added.

The published Double 7 against owning the index, SPY 1993-2026, both scored by the same engine on the same $35,000 basis.

Profit factor, gross winnings divided by gross losses, of 2.33 means the winners paid for the losers more than twice over. The average trade returned 0.67%, the largest single win was $2,182.80 and the largest single loss $2,971.71, and the worst losing streak was four trades in a row.

Now read the two columns against each other, because a lot of Double 7 writing gets this backwards. Buy-and-hold made more than twice the annual return. If your goal is the biggest number at the end, you should have done nothing and owned the index. The strategy's case is entirely in the other rows: it was in cash more than two-thirds of the time, and its worst valley was 8.5% deep against the index's 56.5%. Scored identically through the same metric engine, which compounds each trade's own percentage return, annualises that at the strategy's trade frequency and divides by the worst percentage drawdown, that is 0.76 against 0.22, or 3.5 times the return for each unit of pain.

That is the honest shape of this strategy. It is not a way to beat the market. It is a way to be paid while taking a worst fall under a third of the index's.

For comparison, QuantifiedStrategies publishes a Double Seven backtest with 154 trades since 1993, an 82.5% win rate and a 2.58 profit factor. Our trade count is higher and our win rate lower, and the difference this study can name is the one above: they fill at the close, we fill the next morning. Their run was not reconstructed here, so nothing rules out other differences. What this study measures is its own implementation, filled a day later than theirs.

Seven is not the magic number. It is on the plateau.

Why seven days? Is there anything special about the number?

No. And that is the best news here.

I swept the lookback from 2 days to 21, matching the exit length to the entry each time so the rule keeps its shape:

The Double 7 with its lookback swept, exit matched to entry, SPY long, 200-day filter on, no other filters.

The published 7 ranks third of eight. The 5-day version scores 0.99 and the 4-day 0.82. If Connors had called it the Double Five it would have looked slightly better.

The 7 is what got copied, though. The most-used open-source Double 7 script on TradingView hard-codes it, as does nearly every write-up of the rule. That is how a number chosen once in 2008 becomes a number thousands of people trade without ever testing the one beside it.

But nobody should trade the 5 because of this table, and that is the actual finding. Four, five and seven all clear 0.70 on 322 or more trades each. In this SPY sweep the result holds across a band of neighbouring settings, the tested 4-, 5- and 7-day lookbacks, rather than appearing only at 7. A rule that works at exactly one setting and collapses on either side of it is a rule that found something real about your dataset. Seven sits in the middle of a plateau, and the plateau is the evidence, not the peak.

Take the whole grid, not just this table: of the 3,458 SPY variants with at least 50 trades, 66.5% finished profitable, and of the long ones 93.8% did, at a median return per unit of drawdown of 0.18.

Now look at the win-rate column again. It climbs without a single reversal, from 62.1% at two days to 83.9% at twenty-one. If you picked your setting on win rate you would trade the 21-day version, which scores 0.44 against the 5-day's 0.99. Win rate answers how often. It never answers how much. Every version in the lookback table wins most of the time, and the published one loses bigger than it wins. The average win is smaller than the average loss, a payoff ratio of 0.79 at the published setting, so the win rate is not a bonus, it is the mechanism, and it tells you almost nothing about which version to trade.

Is the 7-day-high exit actually a good exit?

I expected the exit to be the clever half of the strategy. It is the ornamental half.

Same entries, a close at a 7-day low above the 200-day average, fifteen different ways out:

Nine of the fifteen exits tested, identical entries throughout. The full fifteen are in the study's facts file.

Two things fall out.

The calendar nearly ties the rule. Holding every trade for exactly ten sessions, with no indicator, no signal and nothing to compute, scored 0.74 with a 7.5% worst drawdown, against the published exit's 0.76 and 8.5%. It made less money, $67,778.65 against $75,616.50, and it did it with a shallower worst loss. Whatever the seven-day-high exit is contributing, it is not much, and it is not risk control.

The exit that actually wins is shorter than seven. Selling at a five-day high scored 0.91, the best of the fifteen, with the smallest drawdown in the table at 6.6%. Every target longer than five days scored below it, from the published 0.76 at seven down to 0.33 at twenty-one, so the published exit holds slightly past its best window. The shortest exits are not the answer either: the two-day target scored 0.16 and a zero-bar hold -0.02. The scores climb to a peak at five days and fall away on both sides of it.

The calendar keeps up, and the events data sits beside that: 86.1% of signals reached the strategy's own target within 60 sessions, and the median time to get there was 7 sessions. Touches inside 10 bars were not counted here, so those are two observations and not a decomposition of one exit into the other.

This is not an artefact of my code. Jeff Swanson's independent EasyLanguage implementation of the same rule reaches the same shape, and his follow-up on improving it reaches for the same short-horizon exits.

The clever exit barely beats a calendar.

Do you actually need the 200-day moving average filter?

This is where I was most wrong going in. I assumed the filter was what made the strategy profitable. It is not. What it measurably does is cut the worst drawdown.

The 200-day filter switched on and off at every lookback, SPY long, exit matched to entry.

At the published setting, turning the filter off makes more money: $84,760.40 against $75,616.50, a gain of $9,143.90 and a quarter-point of annual return. It also more than doubles the worst drawdown, 8.5% to 19.6%, and cuts return per unit of drawdown from 0.76 to 0.36.

That pattern holds across the whole sweep. The filter improves the drawdown at 8 of 8 lookbacks and the risk-adjusted return at 7 of 8, and it never improves net profit at four days or longer. What the sweep measured is the trade-off, not the reason behind it. At the published setting the filter takes 322 trades where dropping it takes 437, and those extra signals come with a worst drawdown of 19.6% against 8.5%. Whether they arrived in bear markets, and whether their dips kept falling, is not measured anywhere in this study.

So the honest statement of rule one is not "the 200-day average makes this work". It is: the 200-day average is a drawdown tool that you pay for in returns. If that sounds like a bad deal, remember that the strategy's entire argument against buy-and-hold is the drawdown column. Give that up at the published setting and the worst drawdown goes from 8.5% to 19.6%, and return per unit of drawdown from 0.76 to 0.36. That is still above buy-and-hold's 0.22, with a drawdown well short of its 56.5%, but it is less than half the filtered rule's score, for a quarter-point more annual return.

The 200-day filter is a drawdown tool, not a profit tool.

Adding more filters does not help. Taking the published rule and trading it only when volatility was rising cut the trade count from 322 to 219 and left return per unit of drawdown at 0.69 against 0.76, with less money made, $62,046.74 against $75,616.50. No test, interval or resampling was run on that difference. Trading only when the 100-day trend was falling left 23 trades, far below our 50-trade reliability floor. Of the eight ways to switch on the two regime filters tested here, a 20-day volatility filter and a 100-day trend filter, none scored above the published rule's 0.76. Filters outside those two were not tested.

Has the Double 7 stopped working since the book was published?

No, but it has faded, and 2008 is the only date that can honestly test it. Everything before the book is the era the rule was discovered in. Everything after is out of sample for it.

The published rule split at its own publication date, SPY.

The average trade fell by 30.1% and the profit factor by 28.3%. That is a real, measurable decay, and the strategy is still clearly profitable on the far side of it, at 181 trades, which is a large enough sample to mean something.

By decade the picture is the same and slightly noisier: 1.21% per trade in 1993-1999, 0.48% in the 2000s, 0.45% in the 2010s, and 0.67% in 2020-2026. The 1990s were the strongest decade tested, by a distance. The last two decades, at 0.45% and 0.67% per trade, sit closer to the 0.57% since the book than to the 1990s' 1.21%.

The measurement layer agrees, though it runs on the same bars and the same engine functions as the grid, so it corroborates rather than tests the result again. Averaged over every signal with overlaps kept, the five-day forward return was 0.811% in 1993-1999 and 0.305% in 2010-2019.

It still works. It works about a third less well.

Is the edge real, or is it just luck?

It clears this random baseline, and not narrowly.

The control is a coin flip matched to this study's own median reliable variant, at 239 requested entries and a 5-bar hold, of which it completed 209 on average. It is not matched to the published rule, which fired 322 trades on a different exit, so entry location is not the only difference between the two. Run over 10 seeds it made $9,146.50 with a spread of $11,499.67, won 56.89% of the time, and scored 0.040 on return per unit of drawdown with a spread of 0.055.

The published rule made $75,616.50 at a 74.5% win rate and scored 0.76. That is more than thirteen of the control's seed spreads above its 0.040. Random entries in a rising market made money on average across the ten seeds, which is worth knowing, though the seed-to-seed spread of $11,499.67 is wider than that average. A positive result is not by itself evidence of anything. This one clears the bar by a distance.

Well clear of a random baseline, and not a buy-and-hold beater.

What does one Double 7 trade actually look like?

Every table above is an average. This is the median trade, not the best one, deliberately, because the best one would teach you a number the strategy does not produce.

This is the median trade, not the best one.

SPY drifts lower for a few sessions inside an uptrend, prints a close that is the lowest of the last seven, and the money goes in at the following open. Seven sessions later a close makes a seven-day high, and the position comes out the next morning, eight sessions after it went in, for 1.06%, or $371.70 on a $35,000 account. No target, no stop, no discretion.

That last figure is worth sitting with. $371.70, at 9.58 trades a year. The Double 7 is not a strategy that changes your life in a quarter. In this backtest, left untouched, the rule's 322 trades added up to $75,616.50 over 33.4 years.

Does the Double 7 survive trading costs?

On SPY, at every cost tested here. Everything above is frictionless. Here it is with a round-trip cost charged as a fraction of the position:

The published rule charged commission and slippage together, as basis points of notional.

A basis point is one hundredth of one percent. The whole edge only disappears somewhere near 67 basis points round trip. What SPY actually costs to trade is not measured anywhere in this study. On a wider instrument, or on a stock rather than an index ETF, that margin shrinks fast, and note that with the filter on the 2-day lookback takes 1,571 trades against the published setting's 322, the 3-day 891 and the 5-day 468, so the shortest of them pays the toll almost five times as often.

Does the Double 7 work on futures?

The family works. The number does not travel.

The same rules on E-mini S&P 500 futures, 2007-2026, one contract, 200-day filter on.

The published 7-day version scores 0.08 on futures. The 4-day version scores 0.56 on nearly twice the trades. Same rule, same market, a lookback two days apart, and a seven-fold difference in risk-adjusted result.

Read that as a warning about the seven rather than about futures. On these two instruments the setting did not carry: 7 sat on the SPY plateau at third of eight, and on ES it scored 0.08 where the 4-day version scored 0.56.

One number in that table is about leverage, not about the rule. The engine sizes futures at one contract against $35,000, which at recent index levels is roughly eleven times the account in notional value. That is why the ES drawdowns are multiples of the SPY ones. Compare the percentages between rows, never the dollars against the SPY table.

What the setup itself does, before any trading rule

Strip out the exits and the position sizing and just measure what happens after every one of the 851 signals, overlaps kept:

Forward return from the entry-bar open, every signal, no trading rule applied.

The edge is not in the first session. Buying the open after a seven-day low and selling that same close loses money on average. The average gain builds over days instead, from +0.082% after one session to +0.612% after ten, and on the exit table the ten-bar calendar scores 0.74 against the clever exit's 0.76.

Three conditioning results are worth knowing, and one of them surprised me:

• The deepest reportable dips paid most, but not in order. Signals that were also the lowest close of the last 50 to 99 days averaged 0.780% over the next five sessions against 0.459% for signals that were only seven-day lows, with the middle bucket at 0.382%, below the shallower one, and reached 1% above entry 96.1% of the time against 81.5%. The deeper 100-to-199-day bucket holds four signals and is flagged below the study's 50-trade floor, so it is not part of the comparison.
• The most oversold signals are not the best ones. Signals with RSI(2) under 5, the deepest readings on Connors' own other indicator, averaged 0.270% over five sessions, against 0.630% for signals between 5 and 10. Past a point, oversold stops meaning stretched and starts meaning broken.
• Four consecutive down closes is the sweet spot. Signals arriving after four down closes averaged 0.804%. After five, the average turns negative, though at 31 signals that cell is below our reliability floor and is directional evidence only.

What about the short side?

Mirrored properly, meaning a close at an N-day high below the 200-day average covered at an N-day low, the short side on this SPY grid is a far weaker rule than the long side.

The shorter mirrors are mildly positive: the 3-day version made $26,356.93 across 299 trades, but with a 22.5% drawdown for a score of 0.09. The published 7-day mirror lost $6,135.34 with a 40.6% drawdown. Across all 2,160 short variants in the grid, the best score on 50 or more trades was 0.22, against 1.26 on the long side.

That gap is measured, not explained. Of those 2,160 variants, 803 finished with a positive net profit and 626 of those cleared the 50-trade floor, so the mirrored short side is not a uniform loser. It is a much weaker version of the same rule on the same bars. What makes it weaker was not tested here, and this is one index on one instrument.

The verdict, and the honest limits

Where the hype is right. The Double 7 is a simple rule that still makes money on the S&P 500, and its 0.76 sits more than thirteen seed spreads above a random control matched to the grid's median variant. That control also differs from the published rule in trade count and exit, so the gap does not isolate entry timing on its own. The 74.5% win rate is genuine. The 8.5% worst drawdown against buy-and-hold's 56.5% is genuine, and it is the actual product. Eighteen years after publication it is still making money: $35,684.88 across the 181 trades since 2008, at a 2.02 profit factor.

Where the hype is wrong. Four things:

1. It does not beat buy-and-hold. It made 3.51% a year against 8.82%. Anyone selling it as an index-beater is selling the wrong column.
2. The win rate is not the edge. It rises monotonically with the lookback while the risk-adjusted result peaks in the middle and falls away at both ends.
3. The 7 is not the best setting tested. Third of eight on SPY, and near-worst on futures. Treat it as one point on a plateau, not as a discovered constant.
4. The exit is nearly decorative. A ten-day calendar gets within 0.02 of it with a shallower drawdown.

The limits, stated plainly.

• Everything except the costs section is frictionless, with no commission and no slippage. The costs section shows the rule profitable at every cost tested on SPY, up to 10 basis points round trip, with the edge estimated to run out near 67, but that margin is instrument-specific and much thinner on anything less liquid.
• One market, two instruments, both S&P 500. SPY and ES are the same underlying index. This is not evidence about other markets, and the futures result is a live demonstration that a setting does not travel.
• No dividends. The buy-and-hold benchmark is price-only. No total-return benchmark was computed here, so every gap in these tables is a price-only gap.
• The futures sizing is aggressive. One contract against $35,000 is roughly eleven times the account. Percentages transfer between the two tables; dollars do not.
• 4,320 variants is a lot of variants. Any single best cell in a grid that size is partly luck. That is why every conclusion here is drawn from a region, the 4-to-7 plateau or the 8-of-8 filter result, and never from the top row.
• A backtest is not a live edge. Thirty-three years of history is a strong argument and not a promise. The decay table is what that looks like when it shows up in the data.

What this means for you

If you want to trade something in this family, the data points at three things, none of which is the headline.

Do not trade the 7 because it is the 7. Trade somewhere in the 4-to-7 band and expect the result to look like the middle of that band rather than the best cell in it. If you are trading E-mini S&P 500 futures rather than the ETF, the ES test points at 4 rather than 7.

Keep the 200-day filter and understand what you bought. Over the full SPY history it cost $9,143.90 of net profit at the published setting. It is the reason the worst drawdown is 8.5% instead of 19.6%, and the drawdown is the entire reason to run this instead of an index fund.

None of the fourth rules tested here earned a place. The volatility filter, the trend filter and the longer exits all cut return per unit of drawdown while cutting the sample. Some of them improved other measures: one longer exit made more money at a better win rate and profit factor, and one filter cut the worst drawdown from 8.5% to 7.6%. None of them lifted return per unit of drawdown, so the rule stays at its three published lines. On SPY, all eight ways of switching on those two regime filters scored below the published rule's 0.76, so each one made the rule longer and its score lower. Keeping a rule the test did not reward is the failure mode I keep coming back to in why most traders fail the robustness test.

If you want the same treatment applied to Connors' other published rules, the R3 strategy rebuilt for index futures and the RSI(2) strategy with a volume filter are here, and Cumulative RSI is the successor Connors himself moved toward. For the same rebuild treatment applied to a trend rule instead of a mean-reversion one, see 40 In, 20 Out. If the 200-day filter result interested you, mastering market regimes is the longer argument for why a trend gate belongs on a mean-reversion system, and the S&P 500 drawdown record since 1870 is what the 56.5% column actually felt like to live through.

The method matters more than this particular rule. Sweep the parameter instead of trusting the published one. Split the sample at a date that means something. Race the clever part against a dumb baseline. Check the result against a coin flip. That sequence is what the Algo Trading Masterclass teaches, and it is why two of the conclusions that a single equity curve supports do not survive it.

Everything here is research, not investment advice. It is a backtest of a published rule on historical data, with the assumptions stated above; it is not a recommendation to buy or sell anything, and past results do not predict future ones.

I write up one of these every couple of weeks: a strategy everyone repeats, tested properly, with the parts that failed left in. StatOasis.com/Overfit