The Edge That Wasn’t

Section 0 — Swing Detection Method Comparison (6 Methods Tested)

All methods assume 1-bar confirmation lookahead. Tested on EUR/JPY M5, 2016 bars, 0.5p spread.

#	Method	Swings (2016 bars)	P&L	Verdict	Why
1	Pure Wilder N=1 h[i]>h[i-1] AND h[i]>h[i+1]	~454	unusable	RULED OUT	Every local noise bump qualifies
2	N=k window h[i] = max(h[i-k..i+k])	fewer	brittle	RULED OUT	Arbitrary k, regime-dependent
3	Alternation only Keep most extreme of same-type run	same as #1 on price	no improvement	RULED OUT	Redundant on ASI; no structural meaning on price alone
4	TopsBots 3-stage N=1 + alternation + exceeding-extremes gate	~65–80	M1: +260p/56%WR M2: +1039p/75%WR	USED	Self-adaptive structural filter
5	Price-confirmation (new_swt) Accept swing only after opposite swing confirmed	fewer, delayed	retroactive	RULED OUT	Misses the entry; labels past, not live
6	Zigzag + pip threshold Running update, require X pip reversal	10p: too many	10p: −274p 25p: −7p 50p: −65p	RULED OUT	All negative P&L; threshold is regime-dependent

Section 1 — Method Definitions

Method	Input	Algorithm	States	Lag	Notes
M1 SB-P	bar h[]/l[]	N=1 → alternation → exceeding-extremes	5: {+2,+1,0,−1,−2}	1 bar	Candle-responsive. Source: TopsBots_Python
M2 SB-A	ASI[i]	identical pipeline on ASI series	5: {+2,+1,0,−1,−2}	1 bar	Noise-filtered. Leads M1 in 72% of pairs

State encoding: +2 = above active HSP (breakout ↑) | +1 = between, last swing = LSP (ascending context) | 0 = startup | −1 = between, last swing = HSP (descending context) | −2 = below active LSP (breakout ↓)

Section 2 — 25 Joint States: Key Findings

Lag analysis: M2 (ASI) leads M1 (price) in 72% of matched signal pairs (104/145 pairs). Median lead time: 30 min (6 M5 bars). M1 leads in only 17/145. Same-bar agreements: 24.

Joint State	Bars	Avg 1-bar pip	WR%	Signal	Zone
P↓×P↓	437	+3.08p	96%	SHORT	High conviction
P↑×P↑	481	+3.05p	96%	LONG	High conviction
P−×P−	92	+2.97p	87%	SHORT	Confirmed breakout
P↑×P+	110	+2.84p	85%	LONG	Confirmed breakout
P+×P↑	29	+2.81p	100%	LONG	Confirmed breakout
P+×P+	108	+2.61p	94%	LONG	Confirmed breakout
P↓×P−	152	+2.39p	92%	SHORT	Confirmed breakout
P↑×P−	38	+0.93p	71%	SHORT	Weak/reversal
P↓×P+	36	+0.90p	75%	LONG	Weak/reversal
P+×P↓	58	−2.37p	3%	AVOID	TOXIC — false breakout
P−×P↑	56	−2.67p	4%	AVOID	TOXIC — false breakout

Section 2b — 25 Joint State Chart

Section 3 — Comparison with HHHL/LHLL/HHLL/LHHL (4-State Swing Regime)

Source: muzero_asi_mc/micro/nano/picco-ppo/asi_swing_regime_analysis.py — applied to ASI series

The HHHL indicator uses the same 3-stage detection pipeline as TopsBots: N=1 local extremes → alternation enforcement (keep most extreme per cluster) → exceeding-extremes gate (filter_by_exceeding_extremes()). The difference is in what happens after detection. Both produce a list of confirmed HSPs and LSPs from the ASI series. Then:

Dimension	HHHL/LHLL/HHLL/LHHL (4-state regime)	SB-Joint M2 (5-state × 5-state)
Detection pipeline	Identical: N=1 + alternation + exceeding-extremes gate (TopsBots 3-stage) applied to ASI series
What you look at	Last 2 confirmed HSPs and last 2 confirmed LSPs, independently higher_highs = recent_hsp > prev_hsp higher_lows = recent_lsp > prev_lsp	Only the single most recent HSP and LSP (the active swing boundaries). Tracks current price position relative to those two levels.
Question answered	PATTERN of change: "Are the last two highs rising or falling? And separately, are the last two lows rising or falling?" → 2×2 matrix = 4 states	CURRENT POSITION: "Is price above, below, or between its most recent active HSP/LSP?" → 5 states (position + directional context)
Output	One-hot 4-channel per bar, rolling update: HHHL (HH+HL) = uptrend LHLL (LH+LL) = downtrend HHLL (HH+LL) = expanding (bullish divergence) LHHL (LH+HL) = contracting (bearish divergence)	5 states: +2 above HSP | +1 between (asc) | 0 startup | −1 between (desc) | −2 below LSP
Latency	Needs 2 confirmed HSPs + 2 confirmed LSPs to produce a regime label. First signal delayed. Once running: updates on every new confirmed swing.	1-bar lookahead. Active boundaries update as soon as first HSP or LSP is confirmed.
Interpretation	Macro structural trend: is this an uptrend (HH+HL), downtrend (LH+LL), expanding volatility (HH+LL), or squeeze (LH+HL)?	Micro position: right now is price above/below/between its active range, and which direction was the last push?
Relationship	Complementary answers to different questions. HHHL = the swing pattern over time (are we trending, ranging, expanding, contracting?) SB-Joint = the current position relative to active boundaries (are we in range or broken out? which direction is the momentum?) Best combined: HHHL=uptrend + SB-Joint=P↑×P↑ → highest-conviction long entry

• HHHL + SB-Joint=P↑×P↑ = strong uptrend AND currently in ascending context → highest conviction long
• HHHL + SB-Joint=P+×P↓ = uptrend pattern but price above HSP while ASI descending → false breakout warning
• LHHL + SB-Joint=P↑×P↑ = squeeze pattern but both series in ascending context → potential range break up

The key hard rule: BLOCK all entries when SB-Joint ∈ {P+×P↓, P−×P↑}. These are the only states with negative expected value. All other states with non-zero composite signal are profitable.

• HHHL + SB-Joint=P↑×P↑ = strong uptrend AND currently in ascending context → highest conviction long
• HHHL + SB-Joint=P+×P↓ = uptrend pattern but price above HSP while ASI descending → false breakout warning
• LHHL + SB-Joint=P↑×P↑ = squeeze pattern but both series in ascending context → potential range break up

The key hard rule: BLOCK all entries when SB-Joint ∈ {P+×P↓, P−×P↑}. These are the only states with negative expected value. All other states with non-zero composite signal are profitable.

Section 4 — Trading Rules (Hard-coded, Not Learned)

Toxic state entry block (hard rule — do not burden NN with learning this):

Condition	Block	Reason
M1_state == +2 AND M2_state == -1	BLOCK LONG	Price breakout without ASI confirmation — false breakout
M1_state == -2 AND M2_state == +1	BLOCK SHORT	Price breakout without ASI confirmation — false breakout
All other state combinations: allow NN to decide. SB-P + SB-A encoded as 5th+6th NEAT inputs: {-1, -0.5, 0, +0.5, +1}

AND-gate baseline performance (entry only when sign(M1)+sign(M2) > 0): 78% WR, +4.05p avg per trade, +344p over 2016 bars at 0.5p spread

Section 5 — P&L Summary

Method	P&L (2016 bars)	Trades	WR	Avg/trade	Spread
M1 SB-P standalone	+260p	144	56%	+1.81p	0.5p
M2 SB-A standalone	+1039p	247	75%	+4.21p	0.5p
AND gate (M1∧M2)	+344p	85	78%	+4.05p	0.5p
All signals except toxic	+428p	130	77%	+3.29p	0.5p