HCS REVERSAL
At 9:30 snapshot, extremes of a 4-filter Herman Confluence Score fade. Score −4 → reach long 166pt (81% touch); score +4 → fade short. Best R:R on the board.
updated 2026-04-22The setup
At 9:30 ET, compute a 4-filter score (±4 range). Extremes of the score mark exhausted one-sided readings that mean-revert. Middle scores are noise.
- Score = −4 (maximum bearish alignment) → fade up: long at snap price, target snap + 150 pt
- Score = +4 (maximum bullish alignment) → fade down: short at snap price, target snap − 150 pt
The 4 filters
From the original Herman 8-filter system, only 4 survived pressure testing:
| # | Filter | Computation |
|---|---|---|
| 1 | VWAP | price vs session VWAP (00:00 → snap) |
| 2 | EMA50 stack | price vs EMA50 on 1m / 5m / 15m / 1h / 4h / D (≥5 of 6 aligned) |
| 3 | Liq sweep | prior 4h: buy-side swept = bear vote · sell-side swept = bull vote |
| 4 | Open | price vs midnight open |
Each filter contributes +1 (bullish) or −1 (bearish). Sum is the HCS score.
Dropped: Trend slope (inverted), HTF Lvl (wrong-sign at 10:30), Flow + Imbalance (collapsed between 9:30 → 10:30).
Backtest
| Score | Signal | n | Touch rate | First-up rate | Edge |
|---|---|---|---|---|---|
| −4 | FADE_UP | 58 | 81.0% | 56.9% | +7.9pt |
| +4 | FADE_DN | 58 | similar | 40.4% | −8.7pt |
Yearly stability (score = −4 fade-up touch rate): 87 / 75 / 66 / 87 / 81 / 100%. 5 of 6 years clean, one weak year (2023).
At 10:30 snapshot the same 4-filter score produces a continuation signal instead — score +3 is BULL_MAGNET (+12pt asymmetry up), score −3 is BEAR_MAGNET (+9pt asymmetry down). Two different behaviors from the same score depending on when it’s taken.
Entry logic
- 9:30 fire: draw line at snap price ±150pt, prob 75, TTL 390 min (full RTH)
- 10:30 fire: draw line at snap price ±150pt, prob 72, TTL 330 min
Why Tier 1 direction (not magnet)
Unlike AMD and PRELON IB (pure touch magnets), HCS REVERSAL does have a close-direction edge. Reversal at extremes is a real directional signal — the score is measuring exhaustion, and exhausted one-sided reads mean-revert. So this line contributes to both close-direction aggregation and touch-based confluence.
The “only the tails predict” pattern
A common confluence-scoring finding: middle scores (−1 to +1) are pure noise. The predictive content is in the ±4 extremes. This is consistent across most confluence systems — the tails matter, the middle doesn’t.
Middle-score cohorts get no line from this model.
Rarity
Roughly 1–2 fires per month per side on the 9:30 snapshot. Adds another ~1/month on the 10:30 snapshot when regime flips. It’s not an every-day signal — but when it fires it’s one of the highest confidence, best R:R setups on the board.
History
- 2026-04-22 — Shipped after 4 backtests (
backtest_herman_confluence.pyv1/v2,backtest_hcs_slim.pyv3,_magnet.py/_930.py). Final 4-filter slim version validated at 81% touch with 5/6-year consistency. - 2026-04-22 — Paired with standalone NT8 strategy
HcsReversal.cs(long-only config).