Skip to the content
TTHINRIDGE Audit your trades Test your strategy free

The shelf · The 50% pullback · Gold

The 50% pullback on GC

11Failedout of 100

Short answer: no, it does not pass. The 50% pullback on GC, tested by Thinridge on 18.8 years of one-minute Gold futures data (2008 to 2026) with costs on every trade, lost $43,487 net for one contract over 2,721 trades; 3,693 of 5,000 coin-flip copies of it (the same trades, the side picked at random) did as well or better on the years before 2023. Verdict: Failed, 11 of 100.

Where it breaks, and what to change ↓

Swings of 5 bars each side on 5-minute bars. After a swing up, the first pullback to 50% of it that closes back above the line buys, with the stop under the swing's low; the mirror after a swing down. Between 9:45 AM and 3:00 PM; out at 3:30 PM. Run on GC, the Gold. Times are New York time.

swings 5 bars each sidepullback to 50%trade on touch and close backwindow 9:45 AM–3:00 PMout 3:30 PMstop past the setup's extreme · no targetbars 5-minfollowlong and short
What these settings mean
  • follow: trades in the direction of the break or signal.
  • long and short: takes both buys and sells.
  • window: a new trade can start only between these two times.
  • out: any trade still open is closed at this time.
  • stop past the setup’s extreme: a stop-loss just beyond the setup’s own high (for a sell) or low (for a buy); no profit target.
  • bars: the setup is read on bars of this many minutes.
  • swings: a swing high (or low) is a bar with this many lower highs (or higher lows) on each side of it.
  • The other settings are the setup’s own; the sentence above says what they do.

Tested as written above: a stop one tick past the extreme the setup names, no target, out at 3:30 PM. A different stop or target is a different strategy, with its own result: Test your version →

Hypothetical backtest: one contract, after costs, not real trading. Simulated results have limits: they are not actual trading, they are made with the benefit of hindsight, and no account is likely to achieve results like these. The standard statement

2,721trades, 18.8 years
-$43,487net after costs; -$16 on the average trade
3,693 / 5,000coin flips that matched or beat it on the years before 2023
226 monthslongest wait for a new equity high

How to read the score. It has to beat 5,000 coin flips on the 15 years before 2023 (2008 to 2022, kept apart from the years since 2023): the same trades with the side picked at random, so they test the side it takes, not when it trades. The years-before-2023 check and the year-by-year check are scored on those years too: with the coin flips, 60 of the 100 points; the other three use all 18. A high score doesn’t mean easy to hold: read the longest wait for a new high. A score is how the setup did against our published checks on 2008 to 2026; it is not a forecast, and what it does next is unknown. How a score is made.

2008 to 2022: 2,219 trades, -$62,539. Since 2023, the years kept apart: 502 trades, $19,052.

Against 5,000 coin flips the written strategy is a coin (3,693 of 5,000 matched or beat it on the years before 2023); whatever you do on top of the written steps is where any edge is, and your fills are how to find out which part.

Why it failed: it scored under half the points on the years before 2023 holding up, year after year, the coin flips, higher costs and a nudge to its settings.

How the 11 is made: clean data 5 of 5, the years before 2023 holding up 0 of 25, year after year 0 of 20, the coin flips 4 of 15, higher costs 0 of 20, a nudge to its settings 2 of 15. Passed is 70 or more with two more checks; Close is 40 to 69, or 70 or more that misses a check; Failed is under 40, or failed automatically (too few trades, or money only since 2023). How to read a report

The 50% pullback on GC: cumulative net after costs
The 50% pullback on GC: how far below its last high it sat, with the longest wait for a new high shaded
The 50% pullback on GC: the strategy against 5,000 coin flipsThe 50% pullback on GC: net by year, 2008 to 2022 and since

What to do with this

Where it breaks

Where the written rule breaks: its trades on Wednesdays lose in both periods, -$18,251 from 2008 to 2022 and -$17,632 since 2023. What holds in both: its trades on Thursdays, $10,684 and $32,107.

Longs2008 to 2022: 1,145 trades, -$36,279since 2023: 267 trades, $289Test longs only →
Shorts2008 to 2022: 1,074 trades, -$26,260since 2023: 235 trades, $18,763Test shorts only →
Mondays2008 to 2022: 430 trades, -$9,137since 2023: 97 trades, $5,793Test it without Mondays →
Tuesdays2008 to 2022: 440 trades, -$27,586since 2023: 99 trades, -$5,796Test it without Tuesdays →
Wednesdays2008 to 2022: 467 trades, -$18,251since 2023: 110 trades, -$17,632Test it without Wednesdays →
Thursdays2008 to 2022: 437 trades, $10,684since 2023: 103 trades, $32,107Test it without Thursdays →
Fridays2008 to 2022: 445 trades, -$18,248since 2023: 93 trades, $4,579Test it without Fridays →
Closed at a set time instead
2008 to 2022since 2023
10:00 AM-$2,633-$2,775
11:00 AM-$17,571-$3,418
12:00 PM-$9,930$14,577
1:00 PM-$21,401$21,460
2:00 PM-$34,751$9,107
3:00 PM-$50,564$23,265
Every trade still open at that time closed there, the ones already out kept as they were.
How it got out (all the years)at its set time 1,963 trades, $379,655; at the stop 758 trades, -$423,143.

Every number above is counted from this test’s own trades. A red edge loses in both periods, a green one makes money in both. Picking the best tile is fitting: test your own version and it is scored as a second try.

Or test fewer days, as we did for this setup (below): only after a wider-than-usual day, or only after a quieter one. Each opens the form with that one change filled in; the run is yours, free, private, by email.

What failed, and what to test next

Per contract, the typical trade was up $240 at its best and down $220 at its worst before it got out, against a stop $460 away.

439 of its 1,568 losing trades (28%) were up at least $236, what its typical winner made, before they turned into losses. Taken the other way, the losing trades had been up that much 28% of the time: it's how this market moves, not something this rule does, so changing the exit moves where a trade ends, not whether its entry has an edge.

None of these is the rule's own: what failed is the entry, and no change to the stop or the target fixes an entry.

How far each trade went for it and against it before it got out, held against the same trades taken the other way (the side the coin flips take). Read from this test’s own trades; nothing here is run or scored.

What we ran next

We ran the next one. The report’s one change for this setup is its main setting 10% either way, to see whether the result is the setting or the setup. Both, scored the same way as a second try. Swings 4 bars each side: Failed, 11 out of 100, 3,077 trades; 2008 to 2022 -$59,023 (2,769 of 5,000 coin flips as good); since 2023 $11,097. Swings 6 bars each side: Failed, 11 out of 100, 2,374 trades; 2008 to 2022 -$44,185 (3,835 of 5,000 coin flips as good); since 2023 $26,644.

We tested fewer days. The setups that came near a pass mostly trade rarely, so we kept this one to about half its days, both ways, with one condition (yesterday’s range against the 20 days before it), scored the same way as a second try. Only after a wider-than-usual day: Failed, 12 out of 100, 1,241 trades; 2008 to 2022 -$38,935 (3,856 of 5,000 coin flips as good); since 2023 $17,076. Only after a quieter-than-usual day: Failed, 12 out of 100, 1,471 trades; 2008 to 2022 -$20,856 (2,393 of 5,000 coin flips as good); since 2023 $2,125.

What to try next. Fewer trades. Of the setups that trade least (up to about 122 a year), 15 of 133 came near a pass; of the rest, 5 of 402. A condition that keeps fewer days is the first thing worth testing. A stock index: 15 of the 20 setups that came near were on ES, NQ, YM or RTY. What the setups that came near share →

Check your own trades

If you trade a version of this, the trade audit holds your own fills against the written rule and shows, in dollars, what your part adds or costs. Audit your trades →

A strategy the form doesn’t know goes by email, about a week.

The same setup on the other markets

Run unchanged, at the same settings, on 11 markets (the shelf’s six and five more; of those five, only a pass or a Close has a page of its own). Each one is its own test: picking the market that did best after the fact is how a backtest flatters itself. ES, NQ, YM and RTY are all stock-index futures and move together, so their agreeing is weaker evidence than it looks.

  • ES Failed 16
  • NQ Failed 33
  • YM Failed 14
  • RTY Failed 26
  • GC Failed 11 this page
  • CL Failed 11
  • ZN Failed 16
  • ZB Failed 11
  • 6E Failed 11
  • 6J Failed 11
  • NG Failed 11

One setup, tested, every week

The next verdict by email: what was tested, its score, and where it breaks. Free; leave with one click. Nothing on this site is held back for it.

Sent through Buttondown, the newsletter service you join yourself, as Privacy says.

The engine's report, in full

One trade a day, the first signal in the window, on one-minute bars, with one tick of slippage each way and commission. The coin flips are run on the years 2008 to 2022, kept apart from the years since 2023; the rest of the score uses all of them. Tested November 16, 2026, on data to October 30, 2026.

==============================================================
THINRIDGE ENGINE REPORT
==============================================================

SCOPE: FirstRate one-minute data, unadjusted; the rule exactly as locked; nobody checked it by hand.

1. THE RULE AS LOCKED
   setup type  : Fibonacci retracement
   as you locked it:
     The 50% pullback, at the settings the form starts with, run again on data to Oct 30, 2026: swing 5; ratio 0.5; trigger hold; window 9:45 AM to 3:00 PM; exit 3:30 PM; bar 5; mode follow; sides Both
     the years from 2023 kept apart, the split every shelf cell is scored on
   as the engine read it:
   exit model  : stop: one tick past the level the setup itself names (its swing, pattern end, zone edge or swept extreme)
               : target: none (time exit)
               : time exit: always kept as the backstop
   filter      : none
   tries       : #1
   bars        : signals on 5-minute bars built from the one-minute series by clock bucket; every fill, exit and cost is still on the one-minute series
   entries     : one trade a day, the first qualifying signal in the window
   history     : 18.8 years, kept apart from 2023 on (as declared)
               : trades dated 2023 or later are kept apart; the coin flips are run on the trades before 2023, the rest of the score on all of them

2. VERDICT
   SCORE   : 11 / 100
   VERDICT : FAILED
   UNDER HALF THEIR POINTS : the years before 2023, year to year, the coin flips, costs, nudging a setting
   THE SIX CHECKS : clean data 5 of 5; the years before 2023 0 of 25; year to year 0 of 20; the coin flips 4 of 15; costs 0 of 20; nudging a setting 2 of 15

3. RESULT
   Trades 2,721   Net -$43,487.20   Per trade -$15.98
   Since 2023     : 502 trades, $19,051.60
   Before 2023   : 2,219 trades, -$62,538.80
   Win rate 42.4%  (1153 of 2,721 trades closed above zero after costs)
   Exits by type: stop 758, time 1963
   How it splits (every trade, one contract after costs):
                 trades            net     average   winners
   Long            1412    -$35,990.40     -$25.49       43%
   Short           1309     -$7,496.80      -$5.73       41%
   Monday           527     -$3,344.40      -$6.35       41%
   Tuesday          539    -$33,381.80     -$61.93       40%
   Wednesday        577    -$35,883.40     -$62.19       40%
   Thursday         540     $42,791.00      $79.24       50%
   Friday           538    -$13,668.60     -$25.41       42%
   A split is something to look at, not a new rule: keeping only the best side or day is a new version,
   and testing it counts as one more try.
   Time in a trade: half the trades last 2 h 31 min or less; the middle half between 1 h 13 min and 3 h 59 min; the longest 5 h 44 min.

4. THE COIN FLIPS (5,000 copies of the rule that pick the side at random)
   Full window  p = 0.2970  (1,485 / 5,000)
   Before 2023  p = 0.7386  (3,693 / 5,000)   <- the one the score uses
   Counting earlier tries, p (try #1) = 0.7386

5. YEAR BY YEAR (each year before 2023 a window)
   Years 15   profitable 4   profit factor before 2023 / since 0.77
   By year before 2023: 2008 -$18,613.40  2009 -$6,531.20  2010 $2,813.00  2011 $8,380.60  2012 -$11,896.80  2013 -$1,758.40  2014 -$1,920.00  2015 $459.00  2016 -$6,714.20  2017 $1,727.40  2018 -$2,652.60  2019 -$5,652.00  2020 -$8,548.80  2021 -$3,680.60  2022 -$7,950.80

6. COST STRESS (re-simulated)
   1 tick -$43,487.20   2 ticks -$97,907.20   3 ticks -$152,327.20
   net kept after costs at 1 tick -194.5%   profit factor at twice the costs 0.83   Breakeven 0.0 ticks/side

7. FRAGILITY (5,000 resamples of the trades, drawn with replacement)
   Full   CI -$115,478.20 .. $31,110.80   P(net>0) 0.1166
   Pre-2023 CI -$109,225.80 .. -$14,385.80   P(net>0) 0.0056
   bad-case drawdown (95th percentile) / actual 1.09x   Top trade n/a   Top-5 n/a

8. NUDGING A SETTING (the main setting moved 10% either way)
   Net kept on the nudge 0%   Turned to a loss: YES   Settings 5

9. COULD YOU HOLD IT (no points)
   Longest stretch without a new equity high: 226 months (ending 2026-10)
   Rolling 12-month windows negative: 70% of 215   Worst 12 months: -$22,868.00 (ending 2009-01)
   Max drawdown $78,510.60   Median year -$2,652.60   Years of median to recoup the drawdown: n/a (median year not positive)

10. THE CODE FILE (offered when all five hold)
   [ ] passed
   [ ] fewer than 1 in 20 coin flips matched it on the years before 2023, counting earlier tries
   [ ] the range of outcomes on the years before 2023 stays above zero
   [ ] it breaks even at 4 ticks of cost a side or more
   [ ] no single trade is over 25% of the net
   => not offered at this result.

11. HOW THE POINTS WERE COUNTED (each measure, then the points it earns)
   II, the years before 2023: steadiness kept on the years before 2023 (Sharpe ratio) -1.91 (per-trade) → 0 points; money on the years before 2023 → 0 points
   III, year to year: years profitable 0.27 → 0 points; profit factor before 2023 / since 0.77 → 0 points; one year's share → 0 points
   IV, the coin flips: coin-flip p 0.7386 (5,000 coin flips) → 0 points; bad-case drawdown / actual 1.09 → 4 points; top trade n/a → 0 points
   V, costs: net kept after costs -1.945 → 0 points; profit factor at twice the costs 0.83 → 0 points
   VI, nudging a setting: net kept on a nudge 0.0 → 0 points; settings counted 5 → 2 points; turned to a loss: yes → 0 points

12. THE AUTOPSY (where the money came from and went; the rule as it ran, no new try)
   Exits: time 1,963, $379,655.40; stop 758, -$423,142.60
   On the books, before 2023 (net -$62,538.80): 10:00 AM -$2,633.00  11:00 AM -$17,571.00  12:00 PM -$9,929.80  1:00 PM -$21,400.60  2:00 PM -$34,750.60  3:00 PM -$50,564.40
   Before 2023 -$62,538.80, since 2023 $19,051.60; largest year before 2023: 2011 (no share of the net: the years before 2023 lost money overall)
   Nudging a setting 10%: 0% of the net kept; turned to a loss: yes
   Flat stretch: 226 months without a new high (ending 2026-10); median year -$2,652.60; years of median to recoup n/a

   Longest wait for a new equity high: 226 months, median year -$2,652.60.
   The one next run, if you want it: the neighboring setting of the primary trigger (10% either way), as try #2. Prediction on file: a plateau under 25% means the neighbor rejects too; this run is to show it. Three such runs at most, then the verdict stands.

13. IF YOU TRADE THIS BY EYE
   Against 5,000 coin flips the written rule is a coin (3693 of 5,000 matched or beat it on the years before 2023); whatever you do on top of the written steps is where any edge is, and your fills are how to find out which part.
   That is what the record says about the rule as written. It is not what next year says.

WHAT THIS CANNOT TELL YOU: the score assumes this rule walked in
once (see tries) and that your development did not touch the
years before 2023. Only you know both. Not investment advice. No
guarantee of future results. Statistical analysis of a mechanical
rule as submitted.
==============================================================

Raw market data: FirstRate Data (firstratedata.com). Statistical analysis of a mechanical strategy as written; not advice.