Box Funnel
Every eligible leg (relLeg ≥ 0.33 of the biggest tracked leg) becomes a box. What happens to it next?
Most boxes never get the chance to react at all — the far-end rule kills them the moment price beats the "0" anchor before any touch of Level 1.
Invalid (Touch Failed)15.6%
What This Means
Of 2,888,792 eligible boxes across 18 years, only 776,214 (26.9%) ever become a tradeable signal. The
far-end rule alone eliminates nearly 6 in 10 boxes before price even reacts — this is a strict,
low-signal-rate system by design, not a loose one. Every box that does go valid has already survived a real
structural test.
Entry Outcomes — 10pt Stop (Base Case)
Stop = 10pts beyond Level 2 (above L2 for short, below L2 for long), 18-year sample, 1m+3m+5m+15m pooled
Target 30pt
38.2%
needs 25% to break even
Target 50pt
26.4%
needs 17% to break even
Target 70pt
19.8%
needs 12.5% to break even
Target 100pt
14.3%
needs 9% to break even
The Bottom Line — Unfiltered
At a 10pt stop, every single target tested clears its own breakeven win rate across the full 18-year sample.
30pt target needs 25% to break even and delivers 38.2%. But the biggest finding on this page isn't
the base case — it's how much better this gets with three simple filters. Scroll down.
Stop × Target Grid — Win Rate
Every combination tested, full 18-year sample, 1m+3m+5m+15m pooled, no filters applied yet
| Stop | Target 30 | Target 50 | Target 70 | Target 100 |
|---|
| 5pt | 30.2% | 20.2% | 14.9% | 10.7% |
| 7pt | 33.6% | 22.8% | 16.9% | 12.2% |
| 10pt | 38.2% | 26.4% | 19.8% | 14.3% |
| 15pt | 44.5% | 31.6% | 24.1% | 17.6% |
Stop × Target Grid — Expectancy
Win rate alone is misleading — this is the same grid in actual points per trade (win = +target, loss = −stop)
Lower win rate does not mean lower profit. The tightest stop paired with a mid-to-large target wins the LEAST often but pays the MOST per trade — it dominates most of the grid. Adding 15m to the pool pulls the peak further out toward tp=100, since 15m keeps gaining expectancy at bigger targets where 1m already peaked.
| Stop | Target 30 | Target 50 | Target 70 | Target 100 |
|---|
| 5pt | +5.56 | +6.11 | +6.20 | +6.24 |
| 7pt | +5.43 | +5.99 | +6.03 | +6.02 |
| 10pt | +5.26 | +5.83 | +5.81 | +5.76 |
| 15pt | +5.02 | +5.54 | +5.48 | +5.30 |
Counter-Intuitive Finding
5pt stop + 100pt target has the lowest win rate in the whole grid (12.4%) but the highest expectancy
(+8.04 pts/trade). Win rate and profitability are not the same thing. This model performs better as a
low-frequency, high-payoff system than a high-win-rate one — every stop size shows the same pattern:
expectancy keeps climbing as the target gets bigger, even as win rate falls.
The Real Lever — Leg Size
Does the SIZE of the swing leg the fib is drawn on predict how well it performs? (stop=5, target=100)
A clean, monotonic climb with no noise or reversal anywhere — bigger structural legs make this model dramatically more reliable. This is the single strongest lever found across every test on this site.
| Leg Size | Sample Size | Win Rate | Avg Pts/Trade |
|---|
| 10–20pt | 182,155 | 5.4% | +0.70 |
| 20–40pt | 215,793 | 7.8% | +3.21 |
| 40–80pt | 180,324 | 10.5% | +6.01 |
| 80–160pt | 113,901 | 15.3% | +11.10 |
| 160pt+ | 71,194 | 26.2% | +22.47 |
Pushing The Leg Threshold Further
With Way1-only + session hours (7AM–3PM NY) stacked on top — how far does the curve go?
The climb never plateaus in the range tested. Even at leg≥600pt — a genuinely rare, major structural move — sample size stays real (567 trades over 18 years, ~31/year) and performance keeps improving.
| Leg Size | Sample Size | Win Rate | Avg Pts/Trade |
|---|
| 160pt+ | 21,287 | 28.7% | +25.12 |
| 200pt+ | 15,000 | 30.9% | +27.40 |
| 250pt+ | 9,633 | 33.7% | +30.36 |
| 300pt+ | 6,336 | 36.3% | +33.10 |
| 350pt+ | 4,243 | 38.8% | +35.73 |
| 400pt+ | 2,789 | 40.3% | +37.32 |
| 500pt+ | 1,237 | 43.2% | +40.41 |
| 600pt+ | 567 | 51.7% | +49.26 |
Best Practical Combo Found
Way1 · leg ≥ 300pt · session hours 7AM–3PM NY · 5pt stop · 100pt target
→ 36.5% win rate, +33.3 pts/trade, 11,887 trades over 18 years (~660/year) across all four
timeframes pooled. Frequent enough to actually trade, with expectancy far beyond anything else tested across
every model on this site. The leg≥600pt tier performs even better (+49.3 pts/trade on 3m/5m alone) but
fires far less often — a real edge, just a much slower one.
Way 1 vs Way 2
Way1 = simple wick-reject at Level 1. Way2 = wick through both levels, closes back between them. (stop=10, target=30, 18yr)
Way1 is the clearly stronger setup — the deeper Way2 touch does not translate into a better outcome, despite "more confirmation" on paper. This holds at every leg-size threshold and every timeframe tested.
By Direction, Timeframe & Session
Checking for the bull-drift bias seen in every other model on this site (stop=10, target=30, 18yr)
Direction
Win rate at 30pt target
Timeframe
Win rate at 30pt target
Session Hours vs. Overnight
7AM–3PM NY vs. everything else, stop=5pt / target=100pt, expectancy in points per trade
Longer Timeframe, Higher Win Rate
Long (40.8%) beats short (36.2%) by 4.6 points at the base 10pt/30pt setting — smaller than the 20+ point
long/short splits seen on the GBT and raw-tap tests, but not negligible either. A clean ladder shows up across
timeframe: 15m (46.5%) beats 5m (41.8%) beats 3m (39.8%) beats 1m (36.0%) — more bars between swings
means less noise, which is also why the leg-size filter works so well. Session hours add a real but modest
boost on top of the leg-size and Way1 filters above. See the full breakdown on the
timeframe comparison page.
Real Sweep Rule vs. The 10pt Stop
The owner file's own invalidation rule: a valid fib dies the instant the reaction candle's extreme is taken out
98.8% of valid boxes eventually get swept per the real indicator's own rule, given enough time — this does not contradict the win-rate numbers above, since most trades resolve (hit target or the 10pt stop) long before the eventual sweep would occur. The real sweep rule is a much looser, no-time-limit invalidation; the 10pt stop is the tighter, tradeable version measured here.