AUTHOR: Tony Mudau

MAE Edge Autopsy — What the Numbers Actually Said

Date: 2026-07-02
Sources: mae_analytics_20260702_052945.xlsx, MT5 closed-trade export (Jun 2–10), live trade_history
Implementation: fx_sessions.py, trade_fingerprint_guard.py, technical + risk + orchestration gates

This note documents a forensic pass over Maximum Adverse/Favorable Excursion (MAE/MFE) analytics — not because the dashboard looked scary, but because half the workbook said "Terrible entry" and we needed to know whether the bot was trading badly or measuring badly.

Spoiler: both, but fixable.

The question we started with

Why do we have so many terrible entries?

The Excel export's Entry Quality column looked like a graveyard. Before changing strategy, we had to answer:

1. Is the label measuring real entry timing failure?
2. Is the data behind MFE/MAE trustworthy?
3. Are specific pair + direction + session combos systematically bleeding edge?

How "Entry Quality" is defined (read this first)

Entry Quality is not about MAE. It does not mean "you lost" or "price went against you."

It comes from Entry Efficiency (R):

Entry Efficiency (R) = MFE (pips) ÷ SL distance (pips)

Implementation: entry_efficiency_r() / entry_efficiency_label() in excursion_tracking.py.

Intuition: A 42-pip stop with only 4 pips of favorable movement is 0.10R → Terrible — even if you eventually scratch out a tiny win.

Workbook snapshot (91 exported rows)

Summary sheet (100 closed trades):

• Win rate: 44%
• Avg MAE: 86.2 pips vs Avg SL: 36.8 pips → stops often inside noise
• 33% of trades had MAE ≥ SL
• Avg Entry Efficiency: 3.1R ← misleading (median in rows was 0.20R)

When the label works and MFE is recorded correctly, it correlates with outcomes. When MFE is missing or corrupted, Terrible becomes a catch-all.

Finding #1 — Two-thirds of "Terrible" is a measurement problem

29 of 44 Terrible rows had MFE = 0.0.

That alone forces Entry Efficiency = 0 → Terrible, regardless of what actually happened.

Smoking guns

These are wins labeled as terrible entries. The metric is right in theory; the excursion tracker never got the data.

Why MFE goes missing

Most closes went through MT5 history sync rather than live tick-by-tick excursion updates. The repair job only backfills when mae_pips = 0, not when mfe_pips = 0 — so asymmetric gaps persist.

Bar backfill landmine

When the trade's OHLCV time window is empty, backfill falls back to the last 32 bars unrelated to the trade:

if window.empty:
    window = df.tail(min(32, len(df)))

That produced absurd rows like:

Takeaway: Treat Summary Avg MAE 86 pips and Avg Entry Eff 3.1R with skepticism on older mt5_history_sync rows. Use median efficiency and eyeball individual trades.

Finding #2 — When the data is real, Terrible means "never developed"

15 Terrible rows had MFE > 0 but still below 0.2R — genuine timing failure.

The clearest repeat offender:

USDCAD buys — wide stop, no follow-through

Same story four times: ~42-pip stop, never more than ~5 pips favorable. The market never confirmed the idea.

Range-position columns did not show chasing (mostly Mid-range / Pullback). Failure happened after entry on the price path.

Finding #3 — EURAUD sells were a structured leak

All 9 Terrible EURAUD rows were sells. Eight lost; one won with MFE = 0.

The duplicate blast (May 4)

Four sells opened within seconds of each other:

Same idea, 4× risk, identical stop-out. The old fingerprint guard keyed on exact entry price — different fills slipped through.

MT5 history (Jun 2–10) — different picture for EURAUD

In the broker export for early June, EURAUD trades were buys (+$2.69, +$2.65). No EURAUD sells in that window. The pair isn't toxic in all directions — the sell setup was the leak.

Fix applied: hard block EURAUD sells via PAIR_BLOCKED_DIRECTIONS in fx_sessions.py.

Finding #4 — MT5 P&L story: stops kill, client exits survive

From 16 MT5 closed trades (Jun 2–10):

Every loss hit the broker stop

Client-managed exits: 11 wins, 1 small loss (-$0.94).
Stop losses: 0 wins, 4 losses.

The agent exits well when it manages the trade. Full SL hits are where edge goes to die — often in thin or wrong sessions.

EURUSD sell stacking (Jun 3)

Three sells opened within minutes (12:02, 12:05, 12:05 UTC), all closed together at 14:33 — +$3.99, +$4.32, +$6.12. Lucky outcome, bad process: stacked exposure on one thesis.

What we changed (Jul 2, 2026)

1. Pair + direction blocks (fx_sessions.py)

PAIR_BLOCKED_DIRECTIONS = {
    "EURAUD": frozenset({"sell"}),
}

PAIR_BLOCKED_DIRECTION_SESSIONS = {
    "AUDUSD": {"sell": frozenset({"tokyo"})},
    "AUDJPY": {"sell": frozenset({"tokyo"})},
}

Gates:

• pair_direction_allows_trade() — unconditional direction blocks
• pair_direction_session_allows_trade() — session-specific direction blocks

Wired in technical_agent (proposal time) and risk_agent (sizing time).

2. Direction slot duplicate guard (trade_fingerprint_guard.py)

Old guard: same symbol + direction + rounded entry price → missed stacks at 1.16185 / 1.16135 / 1.16096.

New guard:

Helper: same_direction_entry_blocked()
Enforced in: _entry_duplication_guard() (orchestration) + execute_trade() (execution agent).

New skip reasons in logs:

• same_symbol_direction_slot_active
• same_symbol_direction_window_limit
• reject_pair_direction_session_blocked

Gate flow (where trades die now)

run_cycle candidate
    │
    ▼
technical_agent
    ├─ direction resolved
    ├─ pair_direction_allows_trade        ← EURAUD sell blocked
    ├─ pair_direction_session_allows    ← AUDUSD/AUDJPY sell @ tokyo blocked
    ├─ pair_session_allows_trade        ← pair avoid sessions
    └─ proposal built
    │
    ▼
orchestration
    └─ _entry_duplication_guard         ← direction slot + window limit
    │
    ▼
risk_agent
    ├─ pair direction / session recheck
    └─ max_same_symbol_direction = 1
    │
    ▼
execution_agent
    └─ try_acquire_direction_slot        ← atomic at broker send

Patterns worth remembering

1. Label ≠ truth when MFE = 0

Before tuning strategy from Excel, filter:

MFE = 0  →  likely data gap, not necessarily bad entry
MFE > 0 and Eff < 0.2R  →  real timing failure

2. Cross pairs + Asia + short side = caution

EURAUD sells, AUDUSD/AUDJPY sells in Tokyo — both legs or commodity FX without London liquidity. The fix is session-aware direction blocking, not abandoning the pair entirely.

3. Stacked entries multiply regret

Duplicate guards must key on direction, not price. One thesis → one ticket.

4. Stop loss vs client close is the real P&L divider

Improve entry timing and avoid putting stops inside session noise — but the MT5 export shows exit path matters as much as entry label.

Open work (not yet implemented)

These came out of the same investigation; track separately:

Files to read

Related notes: SESSION_AWARENESS.md, trade_failure_analysis.md, TRADE_LIFECYCLE.md.

One-line summary

We looked terrible on paper because MFE was often zero, but the live leaks were real: EURAUD sells, Asia AUD shorts, and stacked same-direction entries. The bot now blocks those setups and refuses to stack multiple tickets on one morning idea — measure the rest after excursion data is fixed.