feat(aem): per-bucket MAE_MULT table + shadow logging completeness

AEM stays shadow-only this sprint (V7 drives live exits); changes affect
what AEM *would have done*, logged to CH for future V7→AEM demotion analysis.

adaptive_exit_engine.py:
- Replace single MAE_MULT_TIER1=3.5 with MAE_MULT_BY_BUCKET dict
  (B3→None disables MAE stop, B4→2.0 strict, B6→6.0 wide band)
- evaluate() return dict extended: mae_mult_applied, mae_threshold,
  atr, p_threshold, giveback_k (all params that drove the decision)
- adaptive_exit_shadow schema extended (new Nullable columns added via
  ALTER TABLE IF NOT EXISTS — backward-compat with pre-sprint rows)
- log_shadow() signature extended: v7_action, v7_exit_reason,
  naive_would_have (TP/STOP/MAX_HOLD counterfactual at same instant)

dolphin_actor.py:
- AEM shadow call now passes V7 head-to-head decision and naive
  counterfactual so future retrospective requires no offline replay
- EsoF listener registered on DOLPHIN_FEATURES map (esof_advisor_latest
  key); label fed into engine._current_esof_label before each step_bar
- S6/bucket loaders (_load_s6_size_table, _load_asset_bucket_data) and
  constructor wiring for the new GREEN engine kwargs

Plan refs: Tasks 5, 7, 10 — V7 path untouched, AEM return value is
never gated, CH shadow is best-effort (daemon thread).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
hjnormey
2026-04-22 06:07:46 +02:00
parent af5156f52d
commit 48dcf3fe13
2 changed files with 232 additions and 18 deletions

View File

@@ -35,7 +35,7 @@ import os
import threading
import time
import urllib.request
from typing import Optional
from typing import Dict, Optional
import numpy as np
@@ -45,11 +45,28 @@ from adaptive_exit.continuation_model import ContinuationModelBank, FEATURE_COLS
# ── Config ────────────────────────────────────────────────────────────────────
P_THRESHOLD = 0.40 # P(continuation) below this → consider exit
GIVEBACK_K = 0.50 # MFE giveback fraction
MAE_MULT_TIER1 = 3.5 # vol multiplier for tier-1 stop
MAE_MULT_TIER1 = 3.5 # fallback multiplier when bucket-specific entry missing
MAE_MULT_TIER2 = 7.0
ATR_WINDOW = 20
MIN_ATR = 1e-6
# Per-bucket MAE multipliers — replaces single MAE_MULT_TIER1 in the stop check.
# Shadow-only this sprint (AEM doesn't drive live exits; V7 does), so this shapes
# what AEM *would have done* for data collection — not actual trade outcomes.
# `None` disables the MAE_STOP gate entirely for that bucket (giveback/time still apply).
# B3 — natural winners; shadow shows 5.05.1 MAE peaks before FIXED_TP succeeds
# B4 — gross-negative alpha; cut fast before drawdown compounds
# B6 — extreme-vol assets; wide band or we trip on noise
MAE_MULT_BY_BUCKET: Dict[int, Optional[float]] = {
0: 3.5,
1: 3.0,
2: 3.5,
3: None,
4: 2.0,
5: 4.0,
6: 6.0,
}
_CH_URL = "http://localhost:8123/"
_CH_HEADERS = {"X-ClickHouse-User": "dolphin", "X-ClickHouse-Key": "dolphin_ch_2026"}
@@ -72,7 +89,13 @@ def _ch_insert(row: dict, db: str = _SHADOW_DB) -> None:
def _ensure_shadow_table() -> None:
"""Create shadow table if it doesn't exist."""
"""Create shadow table if it doesn't exist.
Extended schema (2026-04-21, GREEN S6 sprint) captures the full AEM decision
snapshot plus V7 action at the same instant, so a future AEM-vs-V7 demotion
analysis can replay head-to-head without needing to re-simulate AEM.
New columns are nullable — existing rows (pre-sprint) simply have NULL for them.
"""
ddl = (
f"CREATE TABLE IF NOT EXISTS {_SHADOW_DB}.{_SHADOW_TABLE} ("
"ts DateTime64(6, 'UTC'),"
@@ -90,19 +113,43 @@ def _ensure_shadow_table() -> None:
"action LowCardinality(String),"
"exit_reason LowCardinality(String),"
"actual_exit LowCardinality(String),"
"pnl_pct Float32"
"pnl_pct Float32,"
# ── AEM decision params (Nullable to stay backward-compat) ──
"mae_mult_applied Nullable(Float32),"
"mae_threshold Nullable(Float32),"
"atr Nullable(Float32),"
"p_threshold Nullable(Float32),"
"giveback_k Nullable(Float32),"
# ── V7 head-to-head (authoritative path) ──
"v7_action LowCardinality(Nullable(String)),"
"v7_exit_reason LowCardinality(Nullable(String)),"
# ── Naive counterfactual (what the dumb TP/STOP/MAX_HOLD would have done) ──
"naive_would_have LowCardinality(Nullable(String))"
") ENGINE = MergeTree()"
" ORDER BY (ts_day, asset, ts)"
" TTL ts_day + INTERVAL 90 DAY"
)
# For pre-existing tables, add the new columns idempotently. CH treats
# ADD COLUMN IF NOT EXISTS as a no-op when the column is already present.
alters = [
f"ALTER TABLE {_SHADOW_DB}.{_SHADOW_TABLE} ADD COLUMN IF NOT EXISTS mae_mult_applied Nullable(Float32)",
f"ALTER TABLE {_SHADOW_DB}.{_SHADOW_TABLE} ADD COLUMN IF NOT EXISTS mae_threshold Nullable(Float32)",
f"ALTER TABLE {_SHADOW_DB}.{_SHADOW_TABLE} ADD COLUMN IF NOT EXISTS atr Nullable(Float32)",
f"ALTER TABLE {_SHADOW_DB}.{_SHADOW_TABLE} ADD COLUMN IF NOT EXISTS p_threshold Nullable(Float32)",
f"ALTER TABLE {_SHADOW_DB}.{_SHADOW_TABLE} ADD COLUMN IF NOT EXISTS giveback_k Nullable(Float32)",
f"ALTER TABLE {_SHADOW_DB}.{_SHADOW_TABLE} ADD COLUMN IF NOT EXISTS v7_action LowCardinality(Nullable(String))",
f"ALTER TABLE {_SHADOW_DB}.{_SHADOW_TABLE} ADD COLUMN IF NOT EXISTS v7_exit_reason LowCardinality(Nullable(String))",
f"ALTER TABLE {_SHADOW_DB}.{_SHADOW_TABLE} ADD COLUMN IF NOT EXISTS naive_would_have LowCardinality(Nullable(String))",
]
try:
body = ddl.encode()
req = urllib.request.Request(_CH_URL, data=body, method="POST")
for k, v in _CH_HEADERS.items():
req.add_header(k, v)
urllib.request.urlopen(req, timeout=10)
for stmt in (ddl, *alters):
body = stmt.encode()
req = urllib.request.Request(_CH_URL, data=body, method="POST")
for k, v in _CH_HEADERS.items():
req.add_header(k, v)
urllib.request.urlopen(req, timeout=10)
except Exception as e:
print(f"[AdaptiveExitEngine] Warning: could not create shadow table: {e}")
print(f"[AdaptiveExitEngine] Warning: could not create/alter shadow table: {e}")
# ── Per-trade state ───────────────────────────────────────────────────────────
@@ -324,12 +371,14 @@ class AdaptiveExitEngine:
bucket_id=st.bucket_id,
)
# Decision logic
mae_threshold = max(0.005, MAE_MULT_TIER1 * atr)
# Decision logic — per-bucket MAE multiplier. `None` entry disables the
# MAE_STOP gate for that bucket (giveback + time checks still apply).
mae_mult = MAE_MULT_BY_BUCKET.get(st.bucket_id, MAE_MULT_TIER1)
mae_threshold = max(0.005, mae_mult * atr) if mae_mult is not None else None
action = "HOLD"
exit_reason = ""
if st.mae > mae_threshold:
if mae_threshold is not None and st.mae > mae_threshold:
action = "EXIT"
exit_reason = "AE_MAE_STOP"
elif (st.peak_mfe > 0 and st.mfe < GIVEBACK_K * st.peak_mfe
@@ -352,10 +401,31 @@ class AdaptiveExitEngine:
"bucket_id": st.bucket_id,
"vel_div_entry": st.vel_div_entry,
"vel_div_now": vel_div_now,
"mae_mult_applied": mae_mult,
"mae_threshold": mae_threshold,
"atr": atr,
"p_threshold": P_THRESHOLD,
"giveback_k": GIVEBACK_K,
}
def log_shadow(self, shadow: dict, actual_exit: str = "", pnl_pct: float = 0.0) -> None:
"""Async log a shadow decision to ClickHouse."""
def log_shadow(
self,
shadow: dict,
actual_exit: str = "",
pnl_pct: float = 0.0,
v7_action: Optional[str] = None,
v7_exit_reason: Optional[str] = None,
naive_would_have: Optional[str] = None,
) -> None:
"""Async log a shadow decision to ClickHouse.
V7 head-to-head + naive counterfactual are optional but should be passed
from dolphin_actor whenever available — they enable the future AEM-vs-V7
demotion analysis without needing an offline replay.
"""
def _opt(v):
return None if v is None else float(v)
row = {
"ts": int(time.time() * 1e6),
"trade_id": shadow.get("trade_id", ""),
@@ -372,5 +442,15 @@ class AdaptiveExitEngine:
"exit_reason": shadow.get("exit_reason_shadow", ""),
"actual_exit": actual_exit,
"pnl_pct": float(pnl_pct),
# New AEM-decision params (Nullable-capable)
"mae_mult_applied": _opt(shadow.get("mae_mult_applied")),
"mae_threshold": _opt(shadow.get("mae_threshold")),
"atr": _opt(shadow.get("atr")),
"p_threshold": _opt(shadow.get("p_threshold")),
"giveback_k": _opt(shadow.get("giveback_k")),
# Head-to-head
"v7_action": v7_action,
"v7_exit_reason": v7_exit_reason,
"naive_would_have": naive_would_have,
}
threading.Thread(target=_ch_insert, args=(row,), daemon=True).start()