✓
v3 methodology fixes:
- Look-ahead bias fixed — pivot confirmed at i−SWING_BARS, entry at next bar
- 12-month window (365 days) for better sample coverage
- Absolute CVD threshold ($50M) — no denominator drift over time
- Base rate computed from all candles in the dataset
- Regime split: 20-bar close-to-close Δ >2% = trending, ≤2% = ranging
window 2025-08-22 → 2026-08-22
candle 4H
swing ±5 bars (causal)
CVD fade ≥$50M absolute
large-order ≥$50k
win def. ≥1% move in signal dir. within horizon
Win rate by signal × horizon
| Signal |
N |
2H |
4H |
8H ←key |
24H |
Direction |
>5% below base — counter-signal
⚠ N < 30 — insufficient sample
Win rate at 8H (vs base rate 36.3%)
Regime split at 8H horizon
⚠
Results are not valid findings — look-ahead bias present.
- Swing confirmation used future bars i+1→i+5 to confirm bar i as pivot
- Sample only 22 signals over 90 days — inadequate for significance
- CVD % denominator grew over time, suppressing late-window signals
- No base rate computed — win% cannot be benchmarked against random chance
window 90 days
candle 4H
swing ±5 bars (biased)
CVD fade ≥15% of cumulative total
large-order ≥$50k
win def. ≥1% move in signal dir. within horizon
Total signals
22
across 4 signal types
Best @8H
42.9%
BEAR_CVD n=7 — still below estimated base
Worst @8H
0%
BULL_LCVD n=1 — single sample
Verdict
Inconclusive
Bias makes these results unreliable
Win rate by signal × horizon
| Signal | N |
2H | 4H | 8H ←key | 24H |
Direction |
Win rate at 8H (vs estimated base ~55%)
Regime split
Not computed in v2 — see v3