Selection biasOur own build of a viral 'branch-and-bound strategy search' interface
The proven-best filter combination
NQ, RTH, 2018-2026
The best of 195,203 is exactly as good as the best of pure noise.
The rule
144 base trades at prior-day, opening-range and overnight highs and lows (fade or break, 12 stop / target / time brackets), combined with up to three context filters: hour, weekday, gap, volatility, distance from the open, position versus yesterday, range used, speed of approach, touch number.
How it was tested
Every combination with at least 200 trades in 2018-2024 was evaluated, 195,203 in all, so the winner is the proven optimum of the class. Then the identical search ran 100 times on shuffled outcomes, and the winner faced 2025-26 data the search never saw.
Cause of death
The winner made +0.242 R a trade on 213 training trades. The same exhaustive search on shuffled outcomes found a best of +0.240 R at the median and +0.314 R at the 95th percentile. On 2025-26 the winner lost -0.206 R a trade.
The lesson
A certificate of optimality proves the in-sample maximum, which is the most overfit point in the class. Judge a search by the null of the whole search, then by data it never touched.
Evidence: 2 recorded runs · flagsNO_OOSSELECTION_BIAS (195,203 variants)MODERATE_SAMPLE
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 1bded425efa6 | Training 2018-24 | 2018-01-01 → 2024-12-31 | 213 | +0.242 R | – | – | net of costs | NO_OOSSELECTION_BIAS (195,203 variants) |
| 571210352957 | Holdout 2025-26 | 2025-01-01 → 2026-08-06 | 39 | −0.206 R | – | – | net of costs | MODERATE_SAMPLESELECTION_BIAS (195,203 variants) |
Selection biasOur own scan of option greeks, VIX futures, NYSE TICK and indicators
41,872 options, VIX and TICK direction tests
NQ and ES, 2023-2025
A t-statistic of 4.74 that means nothing.
The rule
Every feature, and every pair of features, used as a long or short rule over fixed intraday windows.
How it was tested
Each market's best t-statistic was compared with the best t-statistic the same scan produces on shuffled data.
Cause of death
The best NQ cell reached t = 4.74, normally a 'one in a million' result. But across 20,936 tests on NQ, pure noise produces a best t of 4.18 at the median and 5.10 at the 95th percentile. The winner is the size of the best noise.
The lesson
Significance must scale with the number of tests. After thousands of tests, t = 2 is meaningless and even t = 4.7 can be luck.
Evidence: 1 recorded run · flagsGROSSNO_OOSSELECTION_BIAS (20,936 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 2177280d3dda | NQ direction cells (gross) | 2023-01-03 → 2025-12-12 | 112 | – | – | – | gross (before costs) | GROSSNO_OOSSELECTION_BIAS (20,936 variants) |
Selection biasTextbook rules: opening-range breakouts, Donchian, EMA crosses, RSI(2), Bollinger, VWAP, prior-session levels, NR4, gap rules
5,296 classic indicator strategies since 2023
NQ, ES, gold, crude, 2023-2026
Zero of 5,296 passed. Random entries passed just as often.
The rule
Each rule family with four exit styles, in the NY, London and Asia sessions, micro and full-size contracts.
How it was tested
Judged only on 2023-26, beside a control of random entries pushed through the same exits and costs.
Cause of death
The best rule (an Asia-session Donchian fade on NQ) reached a pooled t of 1.83 over 2023-26, below the bar. The random-entry control produced the same number of passing families: none.
The lesson
Test every rule against random entries with the same exits and costs. Exits and costs shape a result more than most entry rules do.
Evidence: 1 recorded run · flagsSELECTION_BIAS (29,283 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 36f78cce3b1b | Best rule found (NQ, Asia) | 2015-01-01 → 2026-08-06 | 4,825 | – | – | – | net of costs | SELECTION_BIAS (29,283 variants) |
CostsOur own microstructure study
Fading one-second price impulses
NQ full-size, 2023-2026
The edge is real, and exactly the size of the toll.
The rule
Fade any 10-second move of at least four standard deviations; stop at the impulse's size; exit after 120 seconds; one event per five minutes.
How it was tested
One pre-specified rule, no tuning, full-size NQ at a $9 round turn.
Cause of death
Over 6,850 trades it kept +0.047 points a trade after costs (t 0.26, profit factor 1.01). On micro contracts, where costs are a larger share of every point, the same pattern is a loser.
The lesson
Measure an edge in units of cost. A gross edge that is about one round turn is not an edge.
Evidence: 1 recorded run · flagsNO_OOSSELECTION_BIAS (539 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| b1b6ef91e1d1 | NQ full-size, net | 2023-01-01 → 2026-08-06 | 6,850 | +0.047 pts | 50.2% | 1.01 | net of costs | NO_OOSSELECTION_BIAS (539 variants) |
Look-ahead biasA popular YouTube concept
Candle Range Theory (CRT)
NQ, ES, gold, 2016-2026
Our first version peeked into the future.
The rule
A 4-hour candle sweeps the previous candle's high or low and closes back inside; trade toward the other side.
How it was tested
The first version was run on random-walk data as a control, and every trade was paired with its mirror image: the opposite side with the same stop and target distances.
Cause of death
The first engine read a confirmation from the finished 4-hour candle, information that did not exist at entry. It 'worked' on random noise, which exposed it, and all 42 of its runs were retracted. Measured correctly, NQ 2016-20 made +0.014 R a trade gross, while the mirror trades made +0.084 R.
The lesson
Run every new engine on random-walk data first. A strategy that profits on noise means the engine is broken.
Evidence: 1 recorded run · flagsGROSSNO_OOSSELECTION_BIAS (39 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| e780bb573348 | NQ 2016-20, signal vs mirror (gross) | 2016-01-01 → 2020-12-31 | 494 | +0.014 R | 28.3% | 1.02 | gross (before costs) | GROSSNO_OOSSELECTION_BIAS (39 variants) |
PlaceboA Reddit post about an 11-year ES system
Options expected-move levels
ES, 2023-2025
The level is an ordinary price.
The rule
Yesterday's close plus or minus the one-day at-the-money straddle; fade the first touch.
How it was tested
The same rule run at placebo prices near the real level.
Cause of death
Fading the expected-move line lost -0.127 R a trade over 2023-25. The placebo prices lost -0.123 R. The difference is a Welch t of -0.07.
The lesson
Every level needs a placebo: the same rule at nearby prices. If the placebo does as well, the level is decoration.
Evidence: 1 recorded run · flagsNO_OOSSELECTION_BIAS (6 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| f999f5e1ab62 | ES fade, 2023-25 | 2023-01-01 → 2025-12-12 | 304 | −0.127 R | 46.1% | – | net of costs | NO_OOSSELECTION_BIAS (6 variants) |
PlaceboA Reddit claim: 'lost 99% in 2016-20, made 316% in 2022-26'
Plain 15-minute opening-range breakout
NQ, ES, gold, crude, silver
The claim replicates. The edge does not.
The rule
Trade the first 5-minute close beyond the first 15 minutes' range.
How it was tested
We reproduced the post, then compared it with the same breakout of placebo 15-minute ranges.
Cause of death
NQ made +0.034 R a trade since 2022 (t 1.37) at micro costs, and breakouts of placebo ranges did nearly as well (Welch t 0.84). The profit is NQ's post-2022 opening drift, which any range catches. ES lost in both halves.
The lesson
Replicating a backtest proves the arithmetic, not the mechanism. Ask what else would have made the same money.
Evidence: 1 recorded run · flagsNO_OOSSELECTION_BIAS (10 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| ff67fa931da0 | NQ close-confirmed breakout, 2022-26 | 2022-01-01 → 2026-10-02 | 1,189 | +0.034 R | – | – | net of costs | NO_OOSSELECTION_BIAS (10 variants) |
PlaceboA YouTube setup
The '10 AM model'
NQ, ES, gold, crude, 2015-2026
The magic minute is not magic.
The rule
Mark the 10:00 ET open; price traps one way past it, closes back through, and the retest is the entry with a stop beyond the trap.
How it was tested
720 variants at 10:00, and the identical rules anchored at other times of day.
Cause of death
The plain NQ rule averaged -0.070 R a trade across 2020-26 (t -1.68). The same rules anchored at other times averaged -0.010 R. Nothing about 10:00 stands out.
The lesson
If a setup depends on a special time, run it at every other time too.
Evidence: 1 recorded run · flagsSELECTION_BIAS (38,543 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 6aae44d43303 | NQ plain rule | 2015-01-01 → 2026-08-06 | 1,627 | −0.070 R | – | – | net of costs | SELECTION_BIAS (38,543 variants) |
Out-of-sampleA classic calendar anomaly (McConnell & Xu)
Turn-of-the-month long
NQ
Passed in-sample. Dead on the data we had locked away.
The rule
Buy the open and sell the close on the last trading day of the month and the first three of the next.
How it was tested
It passed on 2015-26, so it went to the vault: 2010-14 data set aside before the study began.
Cause of death
On untouched 2010-14 data, turn-of-month days returned -0.024% against -0.022% on all other days: a t of -0.03.
The lesson
Keep a vault of data you never look at. One clean test beats a hundred in-sample tweaks.
Evidence: 1 recorded run · flagsSELECTION_BIAS (2 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 44f872a6d810 | NQ untouched 2010-14 | 2010-11-29 → 2014-11-24 | 140 | – | 50.0% | 0.87 | net of costs | SELECTION_BIAS (2 variants) |
DecayAn NBER working paper (2026)
Month-end rebalancing front-run
ES, 2015-2026
Real in the paper's years. Gone by the time we tested it.
The rule
Position ahead of pension-fund rebalancing flows signalled by the month's stock-versus-bond returns.
How it was tested
Rebuilt from the paper's appendix, then judged on 2023-26.
Cause of death
Close-to-close, the signal earned +16.2 bp gross in 2015-19, so the construction is right. In 2023-26 it earned +5.9 bp (t 0.98), and the tradable form made +0.037 R a trade (t 0.72), below its permutation bar.
The lesson
Published effects decay. Always test the most recent years on their own.
Evidence: 1 recorded run · flagsSELECTION_BIAS (6 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 119112b59a10 | ES, 2023-26 | 2023-01-01 → 2026-07-31 | 167 | +0.037 R | 53.3% | 1.14 | net of costs | SELECTION_BIAS (6 variants) |
Execution artifactA closed-source TradingView indicator advertising 79-85% reversion odds
'Magic Hour' sweep fade
NQ and ES
The claim is true. The trade loses. The better-looking mirror was a fill artifact.
The rule
When price sweeps 25% beyond the 07:00-08:00 ET range, fade it back toward the range.
How it was tested
The claim was checked first, then the trade, then its mirror.
Cause of death
The swept extreme is revisited before 10:00 on about 90% of days, as advertised, but usually after the stop has been hit. The fade lost -0.345 R a trade in 2023-26 (t -7.40). Its mirror looked better until fills were taken from the next minute instead of the signal minute, and then the difference disappeared.
The lesson
A probability without a stop is not a trade. And check that a result survives a stricter fill rule.
Evidence: 1 recorded run · flagsNO_OOSSELECTION_BIAS (8 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| b7b92fc128ad | NQ fade, 2023-26 | 2023-01-01 → 2026-07-15 | 441 | −0.345 R | 43.8% | 0.52 | net of costs | NO_OOSSELECTION_BIAS (8 variants) |
Win-rate illusionA popular style: momentum entry with an exit that locks in profit early
High-win-rate profit-lock exits
NQ, ES, gold, crude
A two-thirds win rate you can get with random entries.
The rule
Breakout or impulse-bar entries on 5- and 15-minute bars, with stops that ratchet to lock small gains.
How it was tested
48 cells, each beside random entries using the identical exit.
Cause of death
With the locking exit, NQ breakouts won 66.8% of trades at a profit factor of 0.99. Random entries with the same exit won 65.6%. The best cell overall made +0.022 R a trade, against a random 95th percentile of +0.019 R.
The lesson
Win rate is mostly a property of the exit. Compare yours with random entries under the same exit.
Evidence: 1 recorded run · flagsNO_OOSSELECTION_BIAS (48 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 022f29b4654c | Best cell (NQ 15-minute breakout) | 2022-01-01 → 2026-10-02 | 1,668 | +0.022 R | 49.3% | – | net of costs | NO_OOSSELECTION_BIAS (48 variants) |
No edgeA trading-book classic
Holy Grail (Linda Raschke)
NQ, ES, gold, crude, silver, 2022-2026
Zero of 15: negative on every market and timeframe since 2022.
The rule
ADX(14) above 30, a pullback to the 20 EMA, a stop entry beyond the touch bar.
How it was tested
5-, 15- and 60-minute bars on five markets, against random-entry percentiles.
Cause of death
Every cell lost money from 2022. On 5-minute bars it lost more than its costs, e.g. NQ -0.093 R a trade (t -2.07).
The lesson
A setup's fame is not evidence. The best-known rules are the most arbitraged.
Evidence: 1 recorded run · flagsNO_OOSSELECTION_BIAS (15 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 352b7c99f167 | Five markets, three timeframes | 2022-01-01 → 2026-10-02 | 5,951 | – | – | – | net of costs | NO_OOSSELECTION_BIAS (15 variants) |
Filter backfiresA trading-book classic
Momentum Pinball (Raschke & Connors)
NQ, ES, gold, crude, silver, 2022-2026
The filter picks exactly the days when breakouts fail.
The rule
When RSI(3) of the daily change is below 30, buy a break of the next day's first-hour high (mirror above 70).
How it was tested
Compared with the same first-hour break taken without the RSI filter.
Cause of death
Without the filter the NQ first-hour break made +0.059 R a trade; with it, -0.102 R (Welch t -2.25). ES: +0.042 R without, -0.150 R with.
The lesson
Always compare a filtered rule with its unfiltered self. A filter can select against you.
Evidence: 1 recorded run · flagsNO_OOSSELECTION_BIAS (5 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 6b217b81f6c1 | Five markets, 2022-26 | 2022-01-01 → 2026-10-02 | 762 | – | – | – | net of costs | NO_OOSSELECTION_BIAS (5 variants) |
Out-of-sampleAn SSRN paper (Valenti)
Intraday asymmetry
NQ, 2015-2026
It worked in 2015-20. 2021-26 carried nothing.
The rule
Yesterday's ln(high/open) + ln(low/open): if negative, buy today's open and sell the close; if positive, short.
How it was tested
The history split in two halves, each judged on its own.
Cause of death
In 2015-20 it separated good from bad days (Welch t 2.42, profit factor 1.11). In 2021-26 that fell to t 0.52 and a profit factor of 1.003. Every year tested also sits inside the paper's own sample, so even the good half is replication, not new evidence.
The lesson
Split your history. A rule that works in one half is a story about that half.
Evidence: 2 recorded runs · flagsNO_OOSSELECTION_BIAS (6 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 721b54017797 | NQ 2015-20 | 2015-01-01 → 2021-01-01 | 1,407 | – | 50.9% | 1.11 | net of costs | NO_OOSSELECTION_BIAS (6 variants) |
| b49451ea24ea | NQ 2021-26 | 2021-01-01 → 2026-08-07 | 1,353 | – | 49.3% | 1.00 | net of costs | NO_OOSSELECTION_BIAS (6 variants) |
PlaceboAn SSRN paper (Chen), without its LLM news filter
Release-time breakout
NQ, 2015-2026
The release adds nothing.
The rule
After each high-impact US data release, arm stop orders at the 200-minute high and low, one cancels the other.
How it was tested
The same rule at event-free clock times.
Cause of death
2015-20 lost -0.046 R a trade (t -1.37). The identical rule at clock times with no release lost -0.045 R. The paper's result does not carry over to NQ futures without the part we could not reproduce.
The lesson
Test the event against no event. If the 'edge' exists without the event, it is not about the event.
Evidence: 1 recorded run · flagsNO_OOSSELECTION_BIAS (9 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 2554987da92e | NQ 2015-20 | 2015-01-01 → 2021-01-01 | 508 | −0.046 R | 43.1% | 0.85 | net of costs | NO_OOSSELECTION_BIAS (9 variants) |
No edgeA staple of order-flow trading
Order-flow absorption
ES and NQ
With the true buy and sell side of every print, absorption still predicts nothing.
The rule
A one-minute bar where heavy aggressive selling fails to push price down (or buying fails to lift it) signals a reversal.
How it was tested
Measured with the true aggressor side of every trade from tick recordings, not the usual tick-rule estimate.
Cause of death
On ES the win rate after absorption was 37.5%, against a 38.5% baseline for the same bracket, a net loss of about one point a trade. No cell passed on ES or NQ.
The lesson
Order-flow reads need the same test as any indicator: a baseline and a random control.
Evidence: 1 recorded run · flagsNO_OOSSELECTION_BIAS (32 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 437a52f2178e | ES, true aggressor side | 2025-12-16 → 2026-08-06 | 12,579 | – | 37.5% | 0.30 | net of costs | NO_OOSSELECTION_BIAS (32 variants) |
No edgeA popular options-flow narrative: dealer charm hedging lifts the last hour
The 'charm' power hour
NQ and ES
A great story with no money in it.
The rule
Buy at 15:00 ET and sell at 15:59, every day or only on days when option positioning implies a strong bid.
How it was tested
Compared with every other RTH hour, and with the days where the story predicted no bid.
Cause of death
Every day: NQ lost -0.097 R a trade over 2018-26 (t -2.72), positive in 1 of 8 years. On the strongest 'charm bid' days: -0.115 R, no better than the days with the weakest (Welch t 0.29).
The lesson
A plausible flow story is a hypothesis, not a signal. Measure it.
Evidence: 2 recorded runs · flagsNO_OOSSELECTION_BIAS (12 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 3aaaff245cb9 | NQ, every day | 2015-02-19 → 2026-10-02 | 2,553 | −0.097 R | – | – | net of costs | NO_OOSSELECTION_BIAS (12 variants) |
| 1c1dde0e8b24 | NQ, 'charm bid' days | 2015-05-29 → 2025-11-18 | 815 | −0.115 R | – | – | net of costs | NO_OOSSELECTION_BIAS (12 variants) |
HindsightCommon trading advice
Just trade the trend days
NQ, 2018-2026
Worth $24,593 a year with perfect hindsight. Not forecastable.
The rule
Identify trend days and enter in their direction with one micro contract.
How it was tested
First the oracle (trend days known in advance), then every forecast we could build.
Cause of death
Known in advance, about 69 trend days a year would pay roughly +$355 each from 10:00 with one micro. But no pre-open feature tells a trend day from a merely big day (every shape-only AUC between 0.48 and 0.52). The best real-time detector made +0.183 R a trade over 2018-26 and -0.115 R in 2025-26.
The lesson
Measure the prize with hindsight first, then measure how much of it you can actually forecast.
Evidence: 3 recorded runs · flagsNO_OOSSELECTION_BIAS (5 variants)GROSSSELECTION_BIAS (10 variants)SELECTION_BIAS (54 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| a98d4bf76bbb | Oracle (not tradeable) | 2018-01-02 → 2026-09-30 | 624 | – | – | – | net of costs | NO_OOSSELECTION_BIAS (5 variants) |
| 08988a102250 | Pre-open forecast quality | 2018-01-02 → 2026-09-30 | 2,168 | – | – | – | gross (before costs) | GROSSNO_OOSSELECTION_BIAS (10 variants) |
| 2c76ed05e3cf | Best real-time detector | 2015-03-02 → 2026-09-30 | 766 | +0.183 R | – | – | net of costs | NO_OOSSELECTION_BIAS (54 variants) |
Out-of-sampleOur own model
Machine-learning top and bottom picker
NQ, train 2024-25, test 2025-26
The flexible model memorised its training years.
The rule
Fade every new session high or low, choosing trades with a classifier on 34 inputs: levels, structure, macro bias, options, order flow, RSI.
How it was tested
Trained on 2024-01 to 2025-06, judged on 2025-07 to 2026-09.
Cause of death
Gradient boosting fit the training window far better than a simple model and lost -0.065 R a trade on the test window. The heavily regularised logistic model held up better (+0.084 R, t 1.15), but its last six months were flat.
The lesson
Complexity has to earn its keep out of sample. Regularise, and split by time, never at random.
Evidence: 2 recorded runs · flagsSELECTION_BIAS (4 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 0039584fdcc4 | Gradient boosting, test window | 2025-07-02 → 2026-09-25 | 350 | −0.065 R | 33.7% | 0.91 | net of costs | SELECTION_BIAS (4 variants) |
| 59b4416ddaf5 | Logistic model, test window | 2025-07-07 → 2026-10-01 | 403 | +0.084 R | 38.7% | 1.13 | net of costs | SELECTION_BIAS (4 variants) |
Win-rate illusionA viral TikTok claim
The '99.6% win rate' 8 AM range
NQ, 2015-2026
True statistic. Worthless trade.
The rule
Mark the 08:00-09:00 ET range; one side is almost always taken, so trade for the other side.
How it was tested
The statistic first, then the trades it suggests, each with its mirror.
Cause of death
A one-hour range is small next to a whole day, so price leaves it on almost every day. That says nothing about which side or before which stop. Trading the sweep back to the other side lost -0.198 R a trade in 2023-26, and its mirror lost about the same.
The lesson
Ask what the statistic would be for any random range. Base rates turn 'win rates' into trivia.
Evidence: 2 recorded runs · flagsNO_OOSSELECTION_BIAS (10 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 28d6c63f7fee | Sweep reversal, 2023-26 | 2023-01-01 → 2026-08-04 | 615 | −0.198 R | 25.5% | 0.93 | net of costs | NO_OOSSELECTION_BIAS (10 variants) |
| b3c612705474 | Other-side fade, 2023-26 | 2023-01-01 → 2026-08-06 | 890 | +0.014 R | 51.6% | 1.11 | net of costs | NO_OOSSELECTION_BIAS (10 variants) |
Filter backfiresA classic weekend effect (Abraham & Ikenberry, 1994)
Friday-to-Monday continuation
NQ and ES, 2018-2026
The opposite of a losing rule is usually another losing rule.
The rule
Trade Friday's late move in its own direction at Monday's open.
How it was tested
Run because fading Friday had lost, which seemed to say that following it would win.
Cause of death
Following Friday made +0.015 R a trade on NQ (t 0.22), positive in 2 of 8 years, no different from midweek. Both directions pay the spread and the stop.
The lesson
Costs and stops hit both sides. Flipping a losing rule doubles the costs, not the edge.
Evidence: 1 recorded run · flagsNO_OOSSELECTION_BIAS (45 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| a66476d8d023 | NQ, 2018-26 | 2015-02-23 → 2026-09-28 | 606 | +0.015 R | – | – | net of costs | NO_OOSSELECTION_BIAS (45 variants) |
No edgeAn ICT concept
SMT divergence
NQ against ES, 2015-2026
The unconfirmed extreme does not reverse. In any window.
The rule
When NQ makes a new session high or low that ES does not confirm, fade NQ.
How it was tested
Three separate windows, at micro costs.
Cause of death
The divergence fade lost in all three windows: -0.170 R a trade in 2015-19 (t -5.19), -0.073 R in 2020-22 and -0.083 R in 2023-26.
The lesson
Test the claimed mechanism directly, and in more than one era.
Evidence: 3 recorded runs · flagsNO_OOSSELECTION_BIAS (1,166 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 446caef5af1c | NQ NY 2015-19 | 2015-02-17 → 2019-12-23 | 699 | −0.170 R | 46.6% | 0.67 | net of costs | NO_OOSSELECTION_BIAS (1,166 variants) |
| 0fc1b87c65ad | NQ NY 2020-22 | 2020-01-02 → 2022-12-30 | 465 | −0.073 R | 47.3% | 0.84 | net of costs | NO_OOSSELECTION_BIAS (1,166 variants) |
| 0618de4c97ff | NQ NY 2023-26 | 2023-01-11 → 2026-08-06 | 449 | −0.083 R | 46.3% | 0.82 | net of costs | SELECTION_BIAS (1,166 variants) |
CostsAn ICT setup, moved to lower timeframes
Silver Bullet on second charts
ES and NQ, 2020-2026
Every step to a finer chart lost more.
The rule
The fair-value-gap setup in its time window, on 1-, 5-, 15-, 30- and 60-second bars.
How it was tested
640 cells across timeframes and filters.
Cause of death
Finer bars mean tighter gaps, tighter stops, and a round turn that eats a bigger share of every trade. The ES 15-second cell averaged -0.337 R a trade across periods (t -2.09 over 2020-26).
The lesson
Lower timeframes do not add information. They add costs.
Evidence: 1 recorded run · flagsSELECTION_BIAS (34,223 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 69841165f11f | ES, 15-second bars | 2020-01-02 → 2026-08-06 | 269 | −0.337 R | – | – | net of costs | SELECTION_BIAS (34,223 variants) |
Win-rate illusionThe 'Judas swing' / liquidity-sweep idea
Overnight-range sweep and reversal
NQ and ES, 2020-2026
No information in either direction.
The rule
After 09:30, the first sweep of the overnight high or low; enter on a one-minute close back inside; target the other side.
How it was tested
Each trade compared with its mirror: the same bars, the other direction, the same distances.
Cause of death
NQ made +0.005 R a trade (t 0.06) with a 19% win rate. The mirror lost -0.040 R. On ES both lost (-0.198 R and -0.166 R). The low win rate is just the geometry of a tight stop against a far target.
The lesson
If the mirror trade does as well, the level carries no information.
Evidence: 4 recorded runs · flagsNO_OOSSELECTION_BIAS (549 variants)
Show the runs behind every number
| Run id | What | Period | Trades | Mean a trade | Win rate | Profit factor | Costs | Flags |
|---|
| 2bd71e81a051 | NQ signal | 2020-01-06 → 2026-08-06 | 1,159 | +0.005 R | 19.1% | 1.01 | net of costs | NO_OOSSELECTION_BIAS (549 variants) |
| c94a2c5c8814 | NQ mirror | 2020-01-06 → 2026-08-06 | 1,159 | −0.040 R | 18.6% | 0.96 | net of costs | NO_OOSSELECTION_BIAS (549 variants) |
| eb9838084bcd | ES signal | 2020-01-02 → 2026-08-06 | 1,066 | −0.198 R | 17.4% | 0.81 | net of costs | NO_OOSSELECTION_BIAS (549 variants) |
| 6e41c69acc6c | ES mirror | 2020-01-02 → 2026-08-06 | 1,066 | −0.166 R | 17.9% | 0.84 | net of costs | NO_OOSSELECTION_BIAS (549 variants) |