Files
OpenFin/backend/services/instrument_service.py
2026-07-24 20:22:21 +02:00

1056 lines
44 KiB
Python

"""
Instrument Dashboard Service.
Provides per-instrument snapshots with price data, technical indicators,
regime detection, trend summary, event filtering, and AI narrative.
"""
import json
import os
import re
import logging
import numpy as np
import pandas as pd
from pathlib import Path
from typing import Dict, Any, List, Optional, Tuple
from datetime import datetime, date
def _base_ticker(t: str) -> str:
"""Normalize Yahoo Finance tickers for comparison: EURUSD=X → EURUSD, BZ=F → BZ, ^GSPC → GSPC."""
return re.sub(r'(=X|=F|=RR|-USD|\^)$', '', t.strip().upper())
# Reverse of the common "yfinance/futures ticker" an Instrument Analysis id (curated or
# quick-added) uses vs. the Cockpit Watchlist's own friendly ticker for the same underlying
# — needed because _base_ticker() only strips a SUFFIX (EURUSD=X -> EURUSD works since the
# base "EURUSD" already equals the Watchlist ticker), it can't turn "GC=F" into "GOLD" or
# "^GSPC" into "SP500" (nothing in common to strip; ^ is a prefix, not a suffix, so
# _base_ticker doesn't even touch it despite what its own docstring claims). Mirrors
# frontend/src/pages/Dashboard.tsx's UNDERLYING_ALIASES table (kept in sync manually — a
# small, stable list, not worth sharing across a Python/TS boundary).
_WAVELET_UNDERLYING_ALIASES = {
"^GSPC": "SP500", "^NDX": "NASDAQ", "^DJI": "DOW", "^RUT": "RUSSELL2000",
"GC=F": "GOLD", "SI=F": "SILVER", "HG=F": "COPPER", "PL=F": "PLATINUM",
"CL=F": "CRUDE", "BZ=F": "BRENT", "NG=F": "NATGAS",
"ZW=F": "WHEAT", "ZC=F": "CORN", "ZS=F": "SOYBEANS",
}
def resolve_watchlist_ticker(instrument_id: str) -> str:
"""Best-effort map from an Instrument Analysis id to the Cockpit Watchlist ticker it
represents — for looking up data keyed by the Watchlist ticker (wavelet cache in
particular). Tries an exact match, then the suffix-stripped form (EURUSD=X -> EURUSD),
then the alias table above (GC=F -> GOLD) — each checked against the tickers actually
in the Watchlist right now, not just "is this a known alias", so a curated catalog id
that happens to share a root with an alias but isn't Watchlist-linked doesn't falsely
resolve. Falls back to the suffix-stripped form if nothing matches."""
uid = instrument_id.strip().upper()
base = _base_ticker(uid)
from services.database import get_instruments_watchlist
watchlist_tickers = {r["ticker"].upper() for r in get_instruments_watchlist()}
if uid in watchlist_tickers:
return uid
if base in watchlist_tickers:
return base
alias = _WAVELET_UNDERLYING_ALIASES.get(uid) or _WAVELET_UNDERLYING_ALIASES.get(base)
if alias and alias in watchlist_tickers:
return alias
return base
logger = logging.getLogger(__name__)
# ── Config loading ─────────────────────────────────────────────────────────────
CONFIG_PATH = Path(__file__).parent.parent / "config" / "instruments.json"
_configs: Optional[Dict[str, Any]] = None
# In-memory narrative cache: key = (instrument_id, iso_date) → str
_narrative_cache: Dict[Tuple[str, str], str] = {}
# "Quick add from Watchlist" — instruments_watchlist's asset_class vocabulary (see
# routers/instruments_watchlist.py's _QUOTE_TYPE_TO_ASSET_CLASS) mapped onto instruments.json's
# own category vocabulary (CATEGORY_ORDER in frontend/src/pages/InstrumentDashboard.tsx).
# Approximate on purpose — this only decides which group heading a quick-added instrument
# is filed under in the picker, it doesn't affect chart/Saxo link/drivers.
_ASSET_CLASS_TO_CATEGORY = {
"forex": "fx", "energy": "energy", "indices": "equity_index",
"etfs": "stock", "equities": "stock", "unknown": "stock",
}
# Generic chart/driver defaults for a quick-added instrument — instruments.json entries are
# hand-curated (custom drivers, ai_context, related_assets...); a quick add has none of that
# yet, just enough to render a chart and price. The user can flesh it out later via the
# existing Drivers editor (update_instrument_drivers), same as any other instrument.
_QUICK_ADD_DEFAULTS = {
"chart": {"ma_periods": [20, 50, 200], "bollinger_period": 20, "bollinger_std": 2, "show_volume": True},
"drivers": [], "event_keywords": [],
"related_assets": [], "correlation_instruments": [], "ai_context": "",
"currency": "USD", "description": "",
# NOT [] — _detect_regime() truncates its 5-slot score vector to len(regime_labels)
# then does np.argmax() on it; an empty list truncates to an empty vector and
# np.argmax([]) raises ValueError, an uncaught 500 for every quick-added instrument
# the moment it has enough real bars (len(df) >= 20) to reach that code path — same
# 5 generic labels _detect_regime() itself falls back to when regime_labels is absent.
"regime_labels": ["Bull", "Bear", "Transition", "Volatile", "Consolidation"],
}
def _load_configs() -> None:
"""instruments.json is the static, git-tracked catalog (baked into the Docker image —
reset to its git contents on every deploy rebuild). Saxo link + drivers are
user-editable at runtime, so they're kept in SQLite (services.database's
instrument_overrides table, on the persistent db_data volume) and merged on top here,
rather than written back to the JSON file where a deploy would silently discard them."""
global _configs
with open(CONFIG_PATH, "r", encoding="utf-8") as f:
data = json.load(f)
_configs = {inst["id"]: inst for inst in data["instruments"]}
from services.database import get_instrument_overrides
overrides = get_instrument_overrides()
for uid, override in overrides.items():
if uid not in _configs:
continue
# saxo_quote_symbol: the override row is the sole source of truth once it exists —
# apply it even when None (an explicit unlink from a previously-linked state).
_configs[uid]["saxo_quote_symbol"] = override.get("saxo_quote_symbol")
# drivers: only overwrite the catalog's own default when this instrument actually
# has a saved override (a row can exist purely for its saxo_quote_symbol, with
# drivers_json left NULL — that must NOT blank out the catalog's base drivers).
if override.get("drivers") is not None:
_configs[uid]["drivers"] = override["drivers"]
# "Quick add from Watchlist" — an override row with `name` set but no matching
# instruments.json entry stands in for a whole catalog entry (see
# set_instrument_override_quick_add / quick_add_instrument_from_watchlist below).
for uid, override in overrides.items():
if uid in _configs or not override.get("name"):
continue
_configs[uid] = {
"id": uid,
"name": override["name"],
"yf_ticker": override.get("yf_ticker") or uid,
"category": override.get("category") or "stock",
"saxo_quote_symbol": override.get("saxo_quote_symbol"),
"drivers": override.get("drivers") or [],
**{k: v for k, v in _QUICK_ADD_DEFAULTS.items() if k not in ("drivers",)},
}
logger.info(f"[instrument_service] Loaded {len(_configs)} instrument configs")
def get_all_instruments() -> List[Dict]:
if _configs is None:
_load_configs()
return list(_configs.values())
def get_instrument(instrument_id: str) -> Optional[Dict]:
if _configs is None:
_load_configs()
return _configs.get(instrument_id.upper())
def update_instrument_drivers(instrument_id: str, drivers: List[Dict]) -> None:
"""Persist updated drivers to the SQLite instrument_overrides table (NOT
instruments.json — that file is baked into the Docker image and gets reset to its
git-tracked contents on every deploy rebuild) and refresh in-memory config."""
global _configs
if _configs is None:
_load_configs()
uid = instrument_id.upper()
if uid not in _configs:
raise ValueError(f"Instrument {uid} not found")
from services.database import set_instrument_override_drivers
set_instrument_override_drivers(uid, drivers)
_configs[uid]["drivers"] = drivers
logger.info(f"[instrument_service] Updated drivers for {uid} ({len(drivers)} drivers)")
def update_instrument_saxo_link(instrument_id: str, saxo_symbol: Optional[str]) -> None:
"""Persist the Saxo quote-symbol link to the SQLite instrument_overrides table (NOT
instruments.json — that file is baked into the Docker image and gets reset to its
git-tracked contents on every deploy rebuild, which is why this link kept disappearing)
and refresh in-memory config. saxo_symbol=None clears the link (falls back to yfinance)."""
global _configs
if _configs is None:
_load_configs()
uid = instrument_id.upper()
if uid not in _configs:
raise ValueError(f"Instrument {uid} not found")
from services.database import set_instrument_override_saxo_symbol
set_instrument_override_saxo_symbol(uid, saxo_symbol)
_configs[uid]["saxo_quote_symbol"] = saxo_symbol
logger.info(f"[instrument_service] Updated Saxo link for {uid}: {saxo_symbol}")
def quick_add_instrument_from_watchlist(ticker: str) -> Dict[str, Any]:
"""Bring a Cockpit watchlist instrument (services.database.instruments_watchlist) into
Instrument Analysis without hand-authoring an instruments.json entry — used by the
picker's "Depuis la Watchlist" section for tickers that don't already have a catalog
entry. Its saxo_quote_symbol (already linked in Config -> Instruments Watchlist) is
copied over automatically — no need to re-link it a second time in Instrument Analysis.
Idempotent: if instrument_id already resolves (either a real catalog entry or an
earlier quick add), returns it unchanged rather than overwriting name/category. Also
tries the yfinance "=X" FX suffix (EURUSD -> EURUSD=X) before creating anything, so a
watchlist ticker that already has a curated catalog counterpart reuses it instead of
fragmenting into a second, duplicate id — this check has to live server-side (not just
in the frontend picker) since _configs is the only always-fresh source of truth."""
global _configs
if _configs is None:
_load_configs()
uid = ticker.strip().upper()
if uid in _configs:
return {"id": uid, "created": False}
if f"{uid}=X" in _configs:
return {"id": f"{uid}=X", "created": False}
from services.database import get_instruments_watchlist, set_instrument_override_quick_add
row = next((r for r in get_instruments_watchlist() if r["ticker"].upper() == uid), None)
if row is None:
raise ValueError(f"'{uid}' n'est pas dans la Watchlist du Cockpit")
category = _ASSET_CLASS_TO_CATEGORY.get(row.get("asset_class") or "unknown", "stock")
set_instrument_override_quick_add(
uid, name=row.get("name") or uid, yf_ticker=uid, category=category,
saxo_quote_symbol=row.get("saxo_quote_symbol"),
)
_load_configs()
logger.info(f"[instrument_service] Quick-added {uid} from Watchlist (category={category})")
return {"id": uid, "created": True}
# ── DataFrame helpers ──────────────────────────────────────────────────────────
def _ohlcv_to_df(records: List[Dict]) -> pd.DataFrame:
"""Convert list of OHLCV dicts (with 'date' key) to a DataFrame indexed by date."""
if not records:
return pd.DataFrame()
df = pd.DataFrame(records)
# Normalise date column — may come as ISO string with or without time component
df["date"] = pd.to_datetime(df["date"], utc=True, errors="coerce")
df = df.dropna(subset=["date"])
df = df.sort_values("date")
df = df.set_index("date")
for col in ("open", "high", "low", "close"):
if col in df.columns:
df[col] = pd.to_numeric(df[col], errors="coerce")
if "volume" in df.columns:
df["volume"] = pd.to_numeric(df["volume"], errors="coerce").fillna(0)
return df
def _safe_float(val) -> Optional[float]:
"""Return a Python float or None for NaN/inf values."""
try:
f = float(val)
return f if np.isfinite(f) else None
except (TypeError, ValueError):
return None
def _round(val, decimals: int = 4) -> Optional[float]:
f = _safe_float(val)
return round(f, decimals) if f is not None else None
# ── Technical indicators ───────────────────────────────────────────────────────
def _compute_indicators(df: pd.DataFrame, config: Dict) -> Dict[str, Any]:
"""
Compute moving averages, Bollinger Bands, RSI14, ATR14, volume MA20.
Returns a dict with time-series lists and scalar latest values.
"""
chart_cfg = config.get("chart", {})
ma_periods = chart_cfg.get("ma_periods", [20, 50, 200])
bb_period = chart_cfg.get("bollinger_period", 20)
bb_std = chart_cfg.get("bollinger_std", 2)
result: Dict[str, Any] = {}
if df.empty or "close" not in df.columns:
return result
close = df["close"]
high = df["high"] if "high" in df.columns else close
low = df["low"] if "low" in df.columns else close
volume = df["volume"] if "volume" in df.columns else pd.Series(dtype=float)
# Moving averages — time-series format for charting
for period in ma_periods:
if len(close) >= period:
ma = close.rolling(period).mean()
valid = ma.dropna()
result[f"ma{period}"] = [
{"time": idx.strftime("%Y-%m-%d"), "value": _round(val)}
for idx, val in valid.items()
]
else:
result[f"ma{period}"] = []
# Bollinger Bands
if len(close) >= bb_period:
bb_ma = close.rolling(bb_period).mean()
bb_sigma = close.rolling(bb_period).std()
bb_upper = bb_ma + bb_std * bb_sigma
bb_lower = bb_ma - bb_std * bb_sigma
valid_idx = bb_ma.dropna().index
result["bb_upper"] = [
{"time": idx.strftime("%Y-%m-%d"), "value": _round(bb_upper[idx])}
for idx in valid_idx
]
result["bb_lower"] = [
{"time": idx.strftime("%Y-%m-%d"), "value": _round(bb_lower[idx])}
for idx in valid_idx
]
result["bb_mid"] = [
{"time": idx.strftime("%Y-%m-%d"), "value": _round(bb_ma[idx])}
for idx in valid_idx
]
else:
result["bb_upper"] = []
result["bb_lower"] = []
result["bb_mid"] = []
# Realized volatility (annualized %, rolling 20d stddev of log returns) — a "how
# turbulent has price action actually been" overlay, distinct from the market-implied
# vol computed elsewhere (vol_surface.py) for the options Strategy Builder.
vol_window = chart_cfg.get("volatility_window", 20)
if len(close) >= vol_window + 1:
log_ret = np.log(close / close.shift(1))
realized_vol = log_ret.rolling(vol_window).std() * np.sqrt(252) * 100
valid_vol = realized_vol.dropna()
result["volatility"] = [
{"time": idx.strftime("%Y-%m-%d"), "value": _round(val, 2)}
for idx, val in valid_vol.items()
]
else:
result["volatility"] = []
# RSI 14
if len(close) >= 15:
delta = close.diff()
gain = delta.clip(lower=0)
loss = (-delta).clip(lower=0)
avg_gain = gain.ewm(com=13, adjust=False).mean()
avg_loss = loss.ewm(com=13, adjust=False).mean()
rs = avg_gain / avg_loss.replace(0, np.nan)
rsi = 100 - (100 / (1 + rs))
valid_rsi = rsi.dropna()
result["rsi14"] = [
{"time": idx.strftime("%Y-%m-%d"), "value": _round(val, 2)}
for idx, val in valid_rsi.items()
]
else:
result["rsi14"] = []
# ATR 14
if len(close) >= 15:
prev_close = close.shift(1)
tr = pd.concat([
high - low,
(high - prev_close).abs(),
(low - prev_close).abs(),
], axis=1).max(axis=1)
atr = tr.ewm(span=14, adjust=False).mean()
valid_atr = atr.dropna()
result["atr14"] = [
{"time": idx.strftime("%Y-%m-%d"), "value": _round(val)}
for idx, val in valid_atr.items()
]
else:
result["atr14"] = []
# Volume MA20
if len(volume) >= 20 and volume.sum() > 0:
vol_ma = volume.rolling(20).mean()
valid_vma = vol_ma.dropna()
result["volume_ma20"] = [
{"time": idx.strftime("%Y-%m-%d"), "value": int(val) if np.isfinite(val) else None}
for idx, val in valid_vma.items()
]
else:
result["volume_ma20"] = []
return result
# ── Regime detection ───────────────────────────────────────────────────────────
def _detect_regime(df: pd.DataFrame, config: Dict) -> Dict[str, Any]:
"""
Score-based regime detection mapped to config.regime_labels.
Labels index: 0=bullish, 1=bearish, 2=transition, 3=volatile, 4=late cycle / consolidation.
"""
regime_labels = config.get("regime_labels", ["Bull", "Bear", "Transition", "Volatile", "Consolidation"])
default = {
"current": regime_labels[0] if regime_labels else "Unknown",
"confidence": 0.0,
"scores": {label: 0.0 for label in regime_labels},
"signals": {},
}
if df.empty or len(df) < 20 or "close" not in df.columns:
return default
close = df["close"]
high = df["high"] if "high" in df.columns else close
low = df["low"] if "low" in df.columns else close
# Moving averages
ma50 = close.rolling(50).mean() if len(close) >= 50 else pd.Series(dtype=float)
ma200 = close.rolling(200).mean() if len(close) >= 200 else pd.Series(dtype=float)
current_price = _safe_float(close.iloc[-1])
current_ma50 = _safe_float(ma50.iloc[-1]) if not ma50.empty else None
current_ma200 = _safe_float(ma200.iloc[-1]) if not ma200.empty else None
# MA slopes (% over n bars)
def _slope_pct(series: pd.Series, lookback: int) -> Optional[float]:
s = series.dropna()
if len(s) < lookback + 1:
return None
v_now = _safe_float(s.iloc[-1])
v_old = _safe_float(s.iloc[-lookback - 1])
if v_now is None or v_old is None or v_old == 0:
return None
return (v_now - v_old) / abs(v_old) * 100
ma50_slope = _slope_pct(ma50, 5) if not ma50.empty else None
ma200_slope = _slope_pct(ma200, 20) if not ma200.empty else None
# Momentum 20d
momentum_20d = None
if len(close) >= 21:
p0 = _safe_float(close.iloc[-21])
p1 = _safe_float(close.iloc[-1])
if p0 and p0 != 0:
momentum_20d = (p1 - p0) / abs(p0) * 100
# Distance from MA200
dist_ma200 = None
if current_price and current_ma200 and current_ma200 != 0:
dist_ma200 = (current_price - current_ma200) / abs(current_ma200) * 100
# ATR vs price (volatility ratio)
atr_pct = None
if len(close) >= 15:
prev_close = close.shift(1)
tr = pd.concat([
high - low,
(high - prev_close).abs(),
(low - prev_close).abs(),
], axis=1).max(axis=1)
atr14 = tr.ewm(span=14, adjust=False).mean().iloc[-1]
if current_price and current_price != 0:
atr_pct = _safe_float(atr14) / current_price * 100 if _safe_float(atr14) else None
# MA50 above MA200 flag
ma50_above_ma200 = None
if current_ma50 is not None and current_ma200 is not None:
ma50_above_ma200 = current_ma50 > current_ma200
signals = {
"ma50_above_ma200": ma50_above_ma200,
"ma50_slope_pct": _round(ma50_slope, 3),
"ma200_slope_pct": _round(ma200_slope, 3),
"momentum_20d_pct": _round(momentum_20d, 3),
"dist_ma200_pct": _round(dist_ma200, 3),
"vol_ratio_pct": _round(atr_pct, 3),
}
# ── Score computation ──────────────────────────────────────────────────────
# bull_score aggregates trend-following signals
bull_score = 0.0
if ma50_above_ma200 is True:
bull_score += 0.4
elif ma50_above_ma200 is False:
bull_score -= 0.4
if ma50_slope is not None:
bull_score += 0.2 if ma50_slope > 0 else -0.2
if momentum_20d is not None:
bull_score += 0.2 if momentum_20d > 0 else -0.2
if ma200_slope is not None:
bull_score += 0.1 if ma200_slope > 0 else -0.1
# Extra penalty/boost for distance extremes
if dist_ma200 is not None:
if dist_ma200 > 15:
bull_score += 0.1 # strong uptrend extension
elif dist_ma200 < -15:
bull_score -= 0.1
# Volatile flag: ATR/price > 3%
is_volatile = (atr_pct is not None and atr_pct > 3.0)
# Transition flag: weak slope + weak momentum
is_transition = (
(ma50_slope is not None and abs(ma50_slope) < 0.1) and
(momentum_20d is not None and abs(momentum_20d) < 1.0)
)
# Late-bull flag: strongly extended above MA200 — risk of reversal
is_late_bull = (
bull_score > 0.6 and
dist_ma200 is not None and dist_ma200 > 20
)
# Correction-in-bull flag: price above MA200 but momentum turning
is_correction_in_bull = (
ma50_above_ma200 is True and
momentum_20d is not None and momentum_20d < -3.0
)
# ── Map to 5 regime slots ──────────────────────────────────────────────────
# Slot 0 = bullish, 1 = bearish, 2 = transition, 3 = volatile, 4 = late/consolidation
raw_scores = [0.0] * 5
if is_correction_in_bull:
# Price still above MA200 but momentum rolling over — partial weight to transition
raw_scores[0] = max(0.0, bull_score * 0.5)
raw_scores[2] = 0.6
elif is_late_bull:
# Extended above MA200: some probability we're in late / consolidation regime
raw_scores[0] = max(0.0, bull_score * 0.55)
raw_scores[4] = bull_score * 0.5
else:
# Bullish
raw_scores[0] = max(0.0, bull_score)
# Bearish
raw_scores[1] = max(0.0, -bull_score)
# Transition (overrides if flagged)
if not is_correction_in_bull:
raw_scores[2] = 0.6 if is_transition else 0.0
# Volatile
raw_scores[3] = 0.7 if is_volatile else 0.0
# Late cycle / consolidation — moderate bull_score but high dist_ma200
if not is_late_bull and not is_correction_in_bull:
if 0.0 < bull_score < 0.3 and dist_ma200 is not None and dist_ma200 > 5:
raw_scores[4] = 0.5
else:
raw_scores[4] = max(0.0, 0.3 - abs(bull_score)) if not is_transition else 0.0
# If volatile, suppress the others a bit
if is_volatile:
raw_scores[0] *= 0.5
raw_scores[1] *= 0.5
raw_scores[2] *= 0.5
# Normalise to sum=1
total = sum(raw_scores)
if total > 0:
norm_scores = [s / total for s in raw_scores]
else:
norm_scores = [1.0 / 5] * 5
# Pad / truncate to match the number of provided labels
n_labels = len(regime_labels)
while len(norm_scores) < n_labels:
norm_scores.append(0.0)
norm_scores = norm_scores[:n_labels]
best_idx = int(np.argmax(norm_scores))
# Cap confidence: max 85% for technical regime (always some model uncertainty)
confidence = _round(min(norm_scores[best_idx], 0.85), 3) or 0.0
scores_dict = {}
for i, label in enumerate(regime_labels):
scores_dict[label] = _round(norm_scores[i], 3) or 0.0
return {
"current": regime_labels[best_idx],
"confidence": confidence,
"scores": scores_dict,
"signals": signals,
}
# ── Trend summary ──────────────────────────────────────────────────────────────
def _get_trend_summary(df: pd.DataFrame) -> Dict[str, Any]:
"""
Return key trend metrics as a flat dict of scalars.
"""
result: Dict[str, Any] = {}
if df.empty or "close" not in df.columns:
return result
close = df["close"]
high = df["high"] if "high" in df.columns else close
low = df["low"] if "low" in df.columns else close
# Current price
result["current_price"] = _round(close.iloc[-1])
# 52-week high / low
n_252 = min(252, len(close))
result["high_52w"] = _round(close.tail(n_252).max())
result["low_52w"] = _round(close.tail(n_252).min())
# MA slopes
def _slope_pct(series: pd.Series, lookback: int) -> Optional[float]:
s = series.dropna()
if len(s) < lookback + 1:
return None
v_now = _safe_float(s.iloc[-1])
v_old = _safe_float(s.iloc[-lookback - 1])
if v_now is None or v_old is None or v_old == 0:
return None
return round((v_now - v_old) / abs(v_old) * 100, 4)
ma50 = close.rolling(50).mean() if len(close) >= 50 else pd.Series(dtype=float)
ma200 = close.rolling(200).mean() if len(close) >= 200 else pd.Series(dtype=float)
result["ma50_slope_5d"] = _slope_pct(ma50, 5)
result["ma200_slope_20d"] = _slope_pct(ma200, 20)
# RSI 14 current
if len(close) >= 15:
delta = close.diff()
gain = delta.clip(lower=0)
loss = (-delta).clip(lower=0)
avg_gain = gain.ewm(com=13, adjust=False).mean()
avg_loss = loss.ewm(com=13, adjust=False).mean()
rs = avg_gain / avg_loss.replace(0, np.nan)
rsi = 100 - (100 / (1 + rs))
result["rsi14_current"] = _round(rsi.iloc[-1], 2)
else:
result["rsi14_current"] = None
# ATR 14 current and vs 3-month average
if len(close) >= 15:
prev_close = close.shift(1)
tr = pd.concat([
high - low,
(high - prev_close).abs(),
(low - prev_close).abs(),
], axis=1).max(axis=1)
atr_series = tr.ewm(span=14, adjust=False).mean()
atr_current = _safe_float(atr_series.iloc[-1])
result["atr14_current"] = _round(atr_current)
# ATR vs 63-day average
if len(atr_series) >= 63:
atr_3m_avg = _safe_float(atr_series.tail(63).mean())
if atr_3m_avg and atr_3m_avg != 0:
result["atr_vs_3m_avg_pct"] = _round((atr_current - atr_3m_avg) / atr_3m_avg * 100, 2)
else:
result["atr_vs_3m_avg_pct"] = None
else:
result["atr_vs_3m_avg_pct"] = None
else:
result["atr14_current"] = None
result["atr_vs_3m_avg_pct"] = None
# Momentum
def _momentum(lookback: int) -> Optional[float]:
if len(close) <= lookback:
return None
p0 = _safe_float(close.iloc[-lookback - 1])
p1 = _safe_float(close.iloc[-1])
if p0 and p0 != 0:
return _round((p1 - p0) / abs(p0) * 100, 3)
return None
result["momentum_1m_pct"] = _momentum(21)
result["momentum_3m_pct"] = _momentum(63)
# Distance from MAs
current_price = _safe_float(close.iloc[-1])
if not ma50.empty and current_price:
ma50_val = _safe_float(ma50.iloc[-1])
if ma50_val and ma50_val != 0:
result["dist_ma50_pct"] = _round((current_price - ma50_val) / abs(ma50_val) * 100, 3)
else:
result["dist_ma50_pct"] = None
else:
result["dist_ma50_pct"] = None
if not ma200.empty and current_price:
ma200_val = _safe_float(ma200.iloc[-1])
if ma200_val and ma200_val != 0:
result["dist_ma200_pct"] = _round((current_price - ma200_val) / abs(ma200_val) * 100, 3)
else:
result["dist_ma200_pct"] = None
else:
result["dist_ma200_pct"] = None
return result
# ── Event filtering ────────────────────────────────────────────────────────────
def _get_relevant_events(
config: Dict,
from_date: Optional[str] = None,
to_date: Optional[str] = None,
instrument_id: Optional[str] = None,
) -> List[Dict]:
"""
Filter market_events DB rows relevant to the instrument, by date range and keyword/asset match.
Also includes events that have a causal analysis for this instrument (regardless of keywords).
Returns max 30 events sorted by start_date asc.
"""
try:
from services.database import get_conn
from services.causal_graphs import init_tables
conn = get_conn()
init_tables(conn) # ensure causal_event_analyses exists
rows = conn.execute("""
SELECT e.*,
(SELECT GROUP_CONCAT(DISTINCT a.instrument)
FROM causal_event_analyses a
WHERE a.market_event_id = e.id) AS analyzed_instruments,
(SELECT a.template_id FROM causal_event_analyses a WHERE a.market_event_id = e.id ORDER BY a.id DESC LIMIT 1) AS cea_template_id,
(SELECT a.id FROM causal_event_analyses a WHERE a.market_event_id = e.id ORDER BY a.id DESC LIMIT 1) AS analysis_id,
(SELECT a.prediction_json FROM causal_event_analyses a WHERE a.market_event_id = e.id ORDER BY a.id DESC LIMIT 1) AS cea_prediction_json,
(SELECT a.actual_json FROM causal_event_analyses a WHERE a.market_event_id = e.id ORDER BY a.id DESC LIMIT 1) AS cea_actual_json,
(SELECT a.activation_score FROM causal_event_analyses a WHERE a.market_event_id = e.id ORDER BY a.id DESC LIMIT 1) AS cea_activation_score
FROM market_events e
ORDER BY e.start_date DESC
""").fetchall()
conn.close()
# Use cea_template_id (from causal_event_analyses) explicitly to avoid
# any collision with an e.* column named template_id
all_events = [
{**dict(r), "template_id": r["cea_template_id"],
"prediction_json": r["cea_prediction_json"],
"actual_json": r["cea_actual_json"],
"activation_score": r["cea_activation_score"]}
for r in rows
]
except Exception as e:
logger.warning(f"[instrument_service] Could not load market events: {e}")
return []
keywords = [kw.lower() for kw in config.get("event_keywords", [])]
related = [ra.lower() for ra in config.get("related_assets", [])]
inst_upper = (instrument_id or "").upper()
filtered = []
for ev in all_events:
ev_start = str(ev.get("start_date", "") or "")
ev_end = str(ev.get("end_date", "") or "")
# Date range overlap filter:
# Include if event overlaps with [from_date, to_date]
# An event overlaps if: ev_start <= to_date AND (ev_end >= from_date OR ev_end is empty)
if to_date and ev_start and ev_start > to_date:
continue
if from_date and ev_end and ev_end < from_date:
continue
# If ev_end is empty (point event), include if start is within [-6 months, to_date]
if from_date and not ev_end:
# Allow events whose start is up to 6 months before the chart window
import datetime
try:
start_dt = datetime.date.fromisoformat(ev_start)
from_dt = datetime.date.fromisoformat(from_date)
if (from_dt - start_dt).days > 180:
continue
except Exception:
pass
# Keyword / asset relevance
ev_name = (ev.get("name") or ev.get("event_name") or "").lower()
ev_desc = (ev.get("description") or "").lower()
ev_assets = (ev.get("affected_assets") or "").lower()
keyword_hit = any(kw in ev_name or kw in ev_desc for kw in keywords)
asset_hit = any(ra in ev_assets for ra in related)
# Always include events that have a causal analysis for this instrument
# Normalize tickers: strip Yahoo Finance suffixes (=X, =F, ^, -USD, =RR) for comparison
analyzed = ev.get("analyzed_instruments") or ""
analyzed_bases = [_base_ticker(s) for s in analyzed.split(",") if s.strip()]
analysis_hit = bool(inst_upper and _base_ticker(inst_upper) in analyzed_bases)
if keyword_hit or asset_hit or analysis_hit:
filtered.append({
"id": ev.get("id"),
"template_id": ev.get("template_id"),
"analyzed_instruments": ev.get("analyzed_instruments"), # comma-sep, e.g. "EURUSD,SP500"
"date": ev_start,
"end_date": ev.get("end_date") or None,
"title": ev.get("name") or ev.get("event_name") or "",
"level": ev.get("level", "medium"),
"category": ev.get("category", ""),
"sub_type": ev.get("sub_type") or None,
"description": (ev.get("description") or "")[:120],
"impact_score": float(ev.get("impact_score") or 0.5),
"expected_value": ev.get("expected_value") or None,
"actual_value": ev.get("actual_value") or None,
"surprise_pct": ev.get("surprise_pct"),
"unit": ev.get("unit") or None,
"absorption_pct": ev.get("absorption_pct"),
"prediction_json": ev.get("prediction_json") or None,
"actual_json": ev.get("actual_json") or None,
"activation_score": ev.get("activation_score"),
})
# Sort by start_date asc for timeline display, cap at 30
filtered.sort(key=lambda e: str(e.get("date", "") or ""))
return filtered[:30]
# ── Main snapshot ──────────────────────────────────────────────────────────────
# Approximate calendar-day span of each yfinance-style period string — used to size the
# Saxo `days` fetch window (Saxo's Chart API is day-count based, not period-string based).
# Same values as routers/wavelet.py's _PERIOD_TO_DAYS; duplicated locally since it's a
# tiny, stable lookup table not worth sharing across modules.
_PERIOD_TO_DAYS = {
"5d": 5, "1mo": 30, "3mo": 90, "6mo": 182,
"1y": 365, "2y": 730, "5y": 1825, "10y": 3650, "max": 3650,
}
def _fetch_ohlcv(config: Dict[str, Any], instrument_id: str, period: str, interval: str) -> Tuple[List[Dict], str, Optional[str]]:
"""Fetch OHLCV records for the snapshot — Saxo-first when the instrument has a
saxo_quote_symbol linked, yfinance otherwise (or as a silent fallback on any Saxo
failure). Returns (records, source, error) — error is the Saxo failure reason, kept
even when the yfinance fallback also comes up empty (a quick-added instrument's
yf_ticker is just its Cockpit ticker, e.g. "BRENT" — that's never a real yfinance
symbol, so surfacing *why Saxo failed* is the only actionable diagnostic in that case,
rather than a bare empty chart with no explanation)."""
saxo_symbol = config.get("saxo_quote_symbol")
saxo_error = None
if saxo_symbol:
from services.database import get_saxo_catalog_by_symbol
entry = get_saxo_catalog_by_symbol(saxo_symbol)
asset_type = entry["asset_type"] if entry else "FxSpot"
try:
from services.saxo_client import get_price_history
days = _PERIOD_TO_DAYS.get(period, 365)
bars = get_price_history(saxo_symbol, asset_type, days=days)
records = [{"date": b["date"], "open": b.get("open"), "high": b.get("high"),
"low": b.get("low"), "close": b.get("close"), "volume": b.get("volume")}
for b in bars]
return records, "saxo", None
except Exception as e:
catalog_note = " — pas dans le catalogue Saxo local, asset_type par défaut" if not entry else ""
saxo_error = f"Saxo ({saxo_symbol}, asset_type={asset_type}{catalog_note}): {e}"
logger.warning(f"[instrument_service] Saxo fetch failed for {instrument_id} ({saxo_symbol}, asset_type={asset_type}): {e}")
yf_ticker = config.get("yf_ticker", instrument_id)
try:
from services.data_fetcher import get_historical
records = get_historical(yf_ticker, period=period, interval=interval)
except Exception as e:
logger.error(f"[instrument_service] Data fetch failed for {instrument_id}: {e}")
records = []
if not records and saxo_error:
return records, "yfinance", saxo_error
return records, "yfinance", None
async def get_snapshot(
instrument_id: str,
period: str = "1y",
interval: str = "1d",
) -> Dict[str, Any]:
"""
Build the full instrument snapshot:
price_data + indicators + regime + trend + events + current price/change.
"""
config = get_instrument(instrument_id)
if not config:
return {"error": f"Unknown instrument: {instrument_id}"}
records, source, source_error = _fetch_ohlcv(config, instrument_id, period, interval)
df = _ohlcv_to_df(records)
# Compute everything
indicators = _compute_indicators(df, config) if not df.empty else {}
regime = _detect_regime(df, config) if not df.empty else {}
trend = _get_trend_summary(df) if not df.empty else {}
# Current price and 1-day change
current_price: Optional[float] = None
change_pct: Optional[float] = None
change_abs: Optional[float] = None
if not df.empty and "close" in df.columns and len(df) >= 2:
last = _safe_float(df["close"].iloc[-1])
prev = _safe_float(df["close"].iloc[-2])
current_price = _round(last)
if last is not None and prev is not None and prev != 0:
change_abs = _round(last - prev)
change_pct = _round((last - prev) / abs(prev) * 100, 3)
# Price data for chart (time + ohlcv)
price_data = []
if records:
for r in records:
price_data.append({
"time": str(r.get("date", ""))[:10],
"open": r.get("open"),
"high": r.get("high"),
"low": r.get("low"),
"close": r.get("close"),
"volume": r.get("volume", 0),
})
# Events spanning the chart period (built after price_data so dates are known)
try:
chart_start = price_data[0]["time"] if price_data else None
chart_end = price_data[-1]["time"] if price_data else None
events = _get_relevant_events(config, from_date=chart_start, to_date=chart_end, instrument_id=instrument_id)
except Exception as e:
logger.warning(f"[instrument_service] Event filtering error: {e}")
events = []
# Macro regime (global cycle: goldilocks / stagflation / recession / etc.)
macro_regime: Dict[str, Any] = {}
try:
from services.data_fetcher import get_macro_gauges, score_macro_scenarios
gauges = get_macro_gauges()
macro_raw = score_macro_scenarios(gauges)
dominant = macro_raw.get("dominant", "incertain")
meta = macro_raw.get("meta", {})
ranked = macro_raw.get("ranked", [])
macro_regime = {
"dominant": dominant,
"label": meta.get(dominant, {}).get("label", dominant) if isinstance(meta, dict) else dominant,
"color": meta.get(dominant, {}).get("color", "#6b7280") if isinstance(meta, dict) else "#6b7280",
"emoji": meta.get(dominant, {}).get("emoji", "🌍") if isinstance(meta, dict) else "🌍",
"scores": macro_raw.get("scores", {}),
"ranked": ranked[:5],
"asset_bias": (macro_raw.get("asset_bias") or {}).get(dominant, {}),
}
except Exception as _me:
logger.warning(f"[instrument_service] Macro regime fetch error: {_me}")
macro_regime = {"dominant": "incertain", "label": "Incertain", "scores": {}}
return {
"instrument": config,
"price_data": price_data,
"indicators": indicators,
"regime": regime,
"macro_regime": macro_regime,
"trend": trend,
"events": events,
"current_price": current_price,
"change_pct": change_pct,
"change_abs": change_abs,
"period": period,
"source": source,
"source_error": source_error if not price_data else None,
}
# ── AI Narrative ───────────────────────────────────────────────────────────────
async def get_narrative(
instrument_id: str,
snapshot_data: Optional[Dict] = None,
force: bool = False,
) -> str:
"""
Generate (or return cached) a 3-4 sentence French narrative for the instrument.
Uses gpt-4o-mini via the existing OpenAI client pattern.
Cache key: (instrument_id, today_iso_date).
"""
today_str = date.today().isoformat()
cache_key = (instrument_id.upper(), today_str)
if not force and cache_key in _narrative_cache:
return _narrative_cache[cache_key]
config = get_instrument(instrument_id)
if not config:
return f"Instrument {instrument_id} non reconnu."
# Fetch snapshot if not provided
if snapshot_data is None:
snapshot_data = await get_snapshot(instrument_id)
trend = snapshot_data.get("trend", {})
regime = snapshot_data.get("regime", {})
drivers = config.get("drivers", [])
ai_ctx = config.get("ai_context", "")
# Build concise prompt data
drivers_str = ", ".join(
f"{d['label']} (poids {d['weight']})" for d in drivers[:5]
)
regime_str = regime.get("current", "N/A")
regime_conf = regime.get("confidence", 0)
rsi = trend.get("rsi14_current")
dist50 = trend.get("dist_ma50_pct")
dist200 = trend.get("dist_ma200_pct")
mom1m = trend.get("momentum_1m_pct")
mom3m = trend.get("momentum_3m_pct")
current_price = trend.get("current_price") or snapshot_data.get("current_price")
change_pct = snapshot_data.get("change_pct")
atr = trend.get("atr14_current")
atr_vs_avg = trend.get("atr_vs_3m_avg_pct")
system_prompt = (
"Tu es un analyste financier quantitatif senior spécialisé en options et macro. "
"Tu génères des narratives d'analyse courtes, précises et actionnables en français."
)
user_prompt = f"""Analyse l'instrument {config['name']} ({instrument_id}) et génère une narrative de 3-4 phrases en français.
## Contexte instrument
{ai_ctx}
## Données techniques actuelles
- Prix actuel : {current_price} | Variation 1j : {change_pct}%
- Momentum 1 mois : {mom1m}% | Momentum 3 mois : {mom3m}%
- Distance MA50 : {dist50}% | Distance MA200 : {dist200}%
- RSI14 : {rsi}
- ATR14 : {atr} | ATR vs moy. 3 mois : {atr_vs_avg}%
## Régime détecté
- Régime : {regime_str} (confiance : {regime_conf:.0%})
## Drivers principaux
{drivers_str}
## Format requis (4 phrases en français, sans bullet points) :
1. Situation technique actuelle (tendance, niveaux clés, momentum)
2. Contexte macro dominant et driver principal
3. Niveaux clés à surveiller (support, résistance, MA critique)
4. Implication pour le trading d'options (vol implicite, stratégie suggérée)"""
# Call OpenAI
try:
import os as _os
from openai import OpenAI
api_key = _os.environ.get("OPENAI_API_KEY", "")
if not api_key:
from services.database import get_config
api_key = get_config("openai_api_key") or ""
if not api_key:
return "Clé API OpenAI non configurée — narrative indisponible."
client = OpenAI(api_key=api_key)
resp = client.chat.completions.create(
model="gpt-4o-mini",
messages=[
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_prompt},
],
temperature=0.3,
max_tokens=400,
)
narrative = resp.choices[0].message.content.strip()
_narrative_cache[cache_key] = narrative
return narrative
except Exception as e:
logger.error(f"[instrument_service] Narrative generation failed for {instrument_id}: {e}")
return f"Narrative indisponible ({type(e).__name__})."