We'd measured monthly Golden Gates on SPX for years. This time we tried to actually trade them, on 17 big-name stocks. The rule that survived is not the one we expected.
Saty's ATR levels in Swing mode: take a stock's monthly average true range (14-period Wilder ATR, always the value from the prior month), anchor it at the prior month's close (PMC), and draw lines at fixed fractions of that range. They reset on the 1st and hold all month.
A downside gate opens when price falls to −38.2% of the monthly ATR below PMC, and completes at −61.8%. The stretch between them is the Golden Gate. The 0 line (PMC itself) is the monthly pivot; a full ATR is −100%. Everything mirrors to the upside.
Every level is known before the month starts — nothing here peeks at the future.
The map, and the one trade on it
Monthly levels for a downside move. The short that survived testing enters at the −61.8% break, risks back to the pivot, and targets the full ATR.
The headline
Gates complete about two-thirds of the time on stocks — but almost none of that probability is money. The one rule that survived every test is a short: sell the −61.8% break in a stock that was still above its 21-day average when the gate opened. It then passed a one-shot, preregistered forward test on 2025–26.
Build sample (2000–2024, 17 stocks): +0.56% per trade after costs, 349 trades, 59.6% winners, profit factor 1.27 — and positive in every era we split it into.
Forward test (2025 → Jul 2026, rules frozen first): +1.19% per trade after costs, 47 trades, 72.3% winners, profit factor 1.65.
It is a modest edge with real caveats (below) — not a money machine. But in a family of studies where almost everything else died, this cell lived through an adversarial audit and an out-of-sample year and a half.
Step 1 · Gates complete a lot — on both sides
How often an opened gate completes by month end
All 17 stocks pooled, 2000–2024, gates with a fair window (≥5 sessions left in the month). Grey tick on each bar group = that side's base rate. "Fast" = price crossed from the trigger to the gate in ≤1 session.
Upside · all
69%
1,505/2,172
Upside · fast
72%
882/1,220
Upside · slow
65%
623/952
Downside · all
67%
1,079/1,603
Downside · fast
69%
680/987
Downside · slow
65%
399/616
upside gatesdownside gatestick = side base rate
The pattern we found on SPX shows up on stocks too — fast traversals complete more often than slow grinds, on both sides. But when we turned "fast" into a trade filter, it added no profit. High completion probability and money are different things. That gap is this whole study.
Step 2 · Why the long side is a mirage
Long trades made money — and so did random longs
Best long rule (buy the +61.8% break, trail out; 1,188 trades) vs random entries in the same stocks held the same lengths of time. Per-trade averages, before costs.
GG long rule
+2.19%
n=1,188
Random longs
+1.97%
matched
These are 2000–2024 mega-caps: they drifted up ~everything. The signal added ≈0.2 points of the 2.2 — statistically indistinguishable from luck.
Simple version: buying big tech on any random day and holding two weeks also made ~2% a trade. The Golden Gate added nothing on the long side. Any long backtest on these names needs this comparison, or it will look brilliant for the wrong reason.
Step 3 · The short that survived
The rule, in plain words
All conditions are known before the fill — this is executable with resting orders.
SetupA downside gate opens (−38.2%) while the stock's previous close is still above its 21-day EMA — a sharp, fresh drop, not an old downtrend.
EntryResting sell-stop at the −61.8% completion level. You only get filled if the drop keeps going.
ExitTarget: −100% (full monthly ATR). Stop: the monthly pivot. Whatever's left closes at month end.
LimitsOne trade per stock per month. Median hold: 4 sessions. About 14 signals/year across 17 names in the build sample.
24 years of the rule, one line
Cumulative percentage points, one unit per trade, 2000–2024 build sample (349 trades). This is the data the rule was found on — the honest test comes next.
Positive in every era: 2000–07 +0.54%/tr · 2008–15 +0.68 · 2016–20 +0.53 · 2021–24 +0.74. 13 of 17 stocks positive; the single biggest loser was shorting NVDA (−51 pp).
The fairest comparison we could build
An audit rebuilt the control to match everything except the signal: random entry days in the same stock and year, same time left in the month, same target and stop distances, same exit engine, even the same above-EMA condition. 10,000 draws.
The GG signal
+0.62%
per trade
Matched random
−0.46%
per trade
Same geometry, no signal → loses money (52% winners, PF 0.84). The signal's +1.07-point gap survived at odds of roughly 3 in 10,000 of being luck. The win rate isn't the trick — the timing is.
Where the money actually comes from
Build sample, 349 trades: how each trade ended, its share of all trades, and its average result.
Hit full-ATR target
53% of trades
avg +5.2%
Stopped at pivot
25%
avg −7.5%
Month-end close
22%
avg −1.3%
All the profit is in the continuation to the full ATR. The leftover month-end holds lose on average — this is not a "hold and hope" trade.
Step 4 · The one-shot forward test
Preregistered · rules frozen first
We wrote the rule down, hashed the file, declared pass/fail lines — then looked at 2025–2026 once.
Pass required: at least 15 trades, positive net average, and ≥50% winners. Result: 47 trades, +1.19% per trade net, 72.3% winners, PF 1.65 — better than the build sample. Excluding the four still-open July 2026 trades it's +0.78%/trade on 43. A pass — read with its caveats, not as a guarantee.
+1.19%Net per trade, forward window47 trades · Jan 2025–Jul 2026
72%Winners out of sample34 of 47 · build was 59.6%
1.65Profit factor forwardbuild was 1.27
The forward test, trade by trade
Cumulative percentage points across the 47 forward trades. Worst single losers: TSLA −15.2%, AMD −11.7% — the pivot stop is wide, so individual losses can be large.
Step 5 · What this trade actually is
It fires on red-tape days
Share of entry days on which the index ETFs closed down, vs any random day in the same period (build sample; the forward window looks the same).
SPY down · entry days
73%
of 349
SPY down · any day
45%
baseline
QQQ down · entry days
76%
of 349
QQQ down · any day
45%
baseline
Despite the "above the 21-day EMA" filter, this is not a bet against strength — by the time the sell-stop fills, the stock has already dropped 0.618 of a monthly ATR, fast. It's a crash-continuation trade, and signals cluster: several names can fire the same day, so they share one risk budget in any real deployment.
Step 6 · If you're long, it's a warning light
For long-only holders
You don't have to short. The earlier signal — a downside gate opening at −38.2% in an uptrending name — reads "reduce or protect," and it's just as informative as waiting for −61.8%.
On 2000–2024 data (473 events): after that signal, the median stock's worst dip within 10 sessions was −4.7%, and a 5%+ drawdown happened 48% of the time. Selling everything for a month was not the answer — by 21 sessions the average name recovered to +1.24%, so a full exit gave up more than it saved on average.
The supported reading: a temporary lighten / hedge / tighter-stop state for a week or two, especially when several of your holdings fire it together — not an automatic liquidation.
Reality checks (we tried to break it)
Three honest dents
Every study on this site reports what didn't survive. These matter for anyone tempted to trade it at size.
Puts: workable, not magicTiming matters enormously. Buying the put at end of day (after the stock fell another ~0.8% past the signal) loses the edge (PF ≈0.9–1.3 at real quotes). Buying at the minute the level trades, at real 1-minute quotes: PF ≈1.8 at mid-market fills, ≈1.1–1.3 if you cross the full spread (32–37 forward trades priced). So puts keep a real chunk of the edge with defined risk — but the "PF 2+" a textbook model promises did not show up in real quotes, and the stock short remains the cleanest expression.
Broker history is flatterRebuilt on a real broker's bars (which only start in 2009), 2010–2024 nets ≈0.0%/trade — a large part of the 24-year edge sits in 2000–2009 and in 2025–26. The forward pass is real, but whether this is durable or partly a recent-regime effect is genuinely open.
Survivor universeThese 17 names are today's winners projected backward. That headwind makes the short results conservative — but it also means the "longs = drift" conclusion is proven only for these survivors, not for all stocks.