Files
OpenFin/backend/services/wavelet_signals.py
2026-07-15 08:47:16 +02:00

384 lines
16 KiB
Python

"""
Automated wavelet signal detection for the watchlist — run once per cycle.
Ported (Python subset) from the trigger-signal detectors in
c:\\DataS\\InstrumentSimulator\\frontend\\src\\main.tsx / frontend/src/lib/waveletTrade.ts.
All 7 trigger kinds from the interactive Wavelets Simulation page are now
available here: extremum, level_threshold, trend_flatten, acceleration,
band_cross, ridge_shift (ssq only), energy_threshold (ssq only).
Parameters (engine + per-signal enable/thresholds) come from the "Technical
Desk" (services.database.get_ai_desk_by_type("technical"), config.signals.wavelet_*)
so they're editable from the existing AI Desks config UI — no hardcoded
defaults here beyond a safe fallback when the desk/key is absent. The
instrument scope stays get_instruments_watchlist() (the desk's own
`instruments` list is NOT used, to avoid reintroducing a second overlapping
instrument-list source).
Every (ticker, band) gets a row every cycle now — signal or not — so the AI
chat context always has fresh slope/energy/ridge state, not just firing
events (see ai_chat_context.py:_block_wavelet_signals).
"""
import json
from typing import Dict, List, Optional
def _compute_slope(series: List[float]) -> List[float]:
n = len(series)
slope = [0.0] * n
for i in range(1, n):
slope[i] = series[i] - series[i - 1]
if n > 1:
slope[0] = slope[1]
return slope
def _compute_acceleration(slope: List[float]) -> List[float]:
n = len(slope)
accel = [0.0] * n
for i in range(1, n):
accel[i] = slope[i] - slope[i - 1]
if n > 1:
accel[0] = accel[1]
return accel
def _avg_slope_range(slope: List[float], frm: int, to: int) -> Optional[float]:
if frm < 0 or to > len(slope) - 1 or to <= frm:
return None
return sum(slope[frm + 1:to + 1]) / (to - frm)
def _build_extremum_signal(series: List[float], direction: str) -> List[bool]:
n = len(series)
raw = [False] * n
for i in range(1, n - 1):
prev_slope = series[i] - series[i - 1]
next_slope = series[i + 1] - series[i]
if direction == "up" and prev_slope > 0 and next_slope <= 0:
raw[i] = True # peak
if direction == "down" and prev_slope < 0 and next_slope >= 0:
raw[i] = True # trough
# raw[j] needs series[j+1] to confirm — shift by one day so the signal never
# requires tomorrow's data.
shifted = [False] * n
for i in range(1, n):
shifted[i] = raw[i - 1]
return shifted
def _build_level_threshold_signal(series: List[float], direction: str, threshold_k: float) -> List[bool]:
n = len(series)
signal = [False] * n
s = 0.0
sq = 0.0
for t in range(n):
s += series[t]
sq += series[t] * series[t]
count = t + 1
if count < 20:
continue # not enough history yet for a stable mean/std
mean = s / count
variance = max(0.0, sq / count - mean * mean)
std = variance ** 0.5
if direction == "up" and series[t] > mean + threshold_k * std:
signal[t] = True
if direction == "down" and series[t] < mean - threshold_k * std:
signal[t] = True
return signal
def _build_trend_flatten_signal(series: List[float], direction: str, trend_days: int, flatten_days: int,
trend_threshold_k: float, flatten_threshold_k: float) -> List[bool]:
n = len(series)
slope = _compute_slope(series)
signal = [False] * n
s = 0.0
sq = 0.0
for t in range(1, n):
s += slope[t]
sq += slope[t] * slope[t]
count = t
if t < trend_days + flatten_days or count < 20:
continue
mean = s / count
variance = max(0.0, sq / count - mean * mean)
std = variance ** 0.5
trend_thresh = trend_threshold_k * std
flatten_thresh = flatten_threshold_k * std
trend = _avg_slope_range(slope, t - flatten_days - trend_days, t - flatten_days)
flat = _avg_slope_range(slope, t - flatten_days, t)
if trend is None or flat is None:
continue
if direction == "up" and trend > trend_thresh and abs(flat) <= flatten_thresh:
signal[t] = True
if direction == "down" and trend < -trend_thresh and abs(flat) <= flatten_thresh:
signal[t] = True
return signal
def _build_acceleration_signal(series: List[float], direction: str, days: int, threshold_k: float) -> List[bool]:
n = len(series)
slope = _compute_slope(series)
accel = _compute_acceleration(slope)
signal = [False] * n
s = 0.0
sq = 0.0
for t in range(2, n):
s += accel[t]
sq += accel[t] * accel[t]
count = t - 1
if t < days or count < 20:
continue
mean = s / count
variance = max(0.0, sq / count - mean * mean)
std = variance ** 0.5
thresh = threshold_k * std
if direction == "up":
if slope[t] <= 0:
continue
ok = True
for d in range(days):
idx = t - d
if idx < 0 or not (accel[idx] < -thresh):
ok = False
break
signal[t] = ok
else:
if slope[t] >= 0:
continue
ok = True
for d in range(days):
idx = t - d
if idx < 0 or not (accel[idx] > thresh):
ok = False
break
signal[t] = ok
return signal
def _build_band_cross_signal(primary: List[float], secondary: List[float], direction: str) -> List[bool]:
n = min(len(primary), len(secondary))
signal = [False] * n
for t in range(1, n):
prev_diff = primary[t - 1] - secondary[t - 1]
curr_diff = primary[t] - secondary[t]
if direction == "down" and prev_diff >= 0 and curr_diff < 0:
signal[t] = True
if direction == "up" and prev_diff <= 0 and curr_diff > 0:
signal[t] = True
return signal
def detect_extremum_signal(series: List[float]) -> Optional[str]:
"""Returns 'up' (confirmed peak) or 'down' (confirmed trough) if the most
recent point is a signal, else None."""
if len(series) < 3:
return None
if _build_extremum_signal(series, "up")[-1]:
return "up"
if _build_extremum_signal(series, "down")[-1]:
return "down"
return None
def detect_level_threshold_signal(series: List[float], threshold_k: float = 2.0) -> Optional[str]:
"""Returns 'up' (overbought) or 'down' (oversold) if the most recent point
breaches a causal z-score threshold, else None."""
if len(series) < 20:
return None
if _build_level_threshold_signal(series, "up", threshold_k)[-1]:
return "up"
if _build_level_threshold_signal(series, "down", threshold_k)[-1]:
return "down"
return None
def detect_trend_flatten_signal(series: List[float], trend_days: int = 10, flatten_days: int = 5,
trend_threshold_k: float = 1.0, flatten_threshold_k: float = 0.3) -> Optional[str]:
if len(series) < trend_days + flatten_days + 20:
return None
if _build_trend_flatten_signal(series, "up", trend_days, flatten_days, trend_threshold_k, flatten_threshold_k)[-1]:
return "up"
if _build_trend_flatten_signal(series, "down", trend_days, flatten_days, trend_threshold_k, flatten_threshold_k)[-1]:
return "down"
return None
def detect_acceleration_signal(series: List[float], accel_days: int = 3, accel_threshold_k: float = 1.5) -> Optional[str]:
if len(series) < accel_days + 20:
return None
if _build_acceleration_signal(series, "up", accel_days, accel_threshold_k)[-1]:
return "up"
if _build_acceleration_signal(series, "down", accel_days, accel_threshold_k)[-1]:
return "down"
return None
def detect_band_cross_signal(primary: List[float], secondary: List[float]) -> Optional[str]:
if len(primary) < 2 or len(secondary) < 2:
return None
if _build_band_cross_signal(primary, secondary, "up")[-1]:
return "up"
if _build_band_cross_signal(primary, secondary, "down")[-1]:
return "down"
return None
def _technical_desk_wavelet_config() -> Dict:
from services.database import get_ai_desk_by_type
desk = get_ai_desk_by_type("technical") or {}
return (desk.get("config") or {}).get("signals") or {}
def scan_watchlist_wavelet_signals() -> List[Dict]:
"""Compute a causal (no-look-ahead) band decomposition for each watchlist
instrument. Every (ticker, band) gets a row every cycle — current slope/
value/energy state always, plus signal_kind/direction/params_json when one
of the enabled trigger kinds fires on the most recent point (first match
wins, evaluated extremum -> level_threshold -> trend_flatten ->
acceleration -> band_cross -> energy_threshold). ridge_shift is evaluated
once per ticker (not per band — the ridge is a single track for the whole
decomposition) and stored as an extra band_label="ridge" row."""
from services.database import get_instruments_watchlist
from services.data_fetcher import get_historical
from services.wavelet_engine import rolling_causal_bands, rolling_causal_bands_ssq
sig_cfg = _technical_desk_wavelet_config()
engine_cfg = sig_cfg.get("wavelet_engine") or {}
if not engine_cfg.get("enabled", True):
return []
num_levels = int(engine_cfg.get("num_levels", 4))
wavelet = engine_cfg.get("wavelet", "gmw")
method = engine_cfg.get("method", "cwt")
lookback = int(engine_cfg.get("lookback_days", 120))
extremum_cfg = sig_cfg.get("wavelet_extremum") or {"enabled": True}
level_cfg = sig_cfg.get("wavelet_level_threshold") or {"enabled": True, "threshold_k": 2.0}
trend_cfg = sig_cfg.get("wavelet_trend_flatten") or {"enabled": False}
accel_cfg = sig_cfg.get("wavelet_acceleration") or {"enabled": False}
cross_cfg = sig_cfg.get("wavelet_band_cross") or {"enabled": False}
ridge_cfg = sig_cfg.get("wavelet_ridge_shift") or {"enabled": False}
energy_cfg = sig_cfg.get("wavelet_energy_threshold") or {"enabled": False}
results: List[Dict] = []
decomposer = rolling_causal_bands_ssq if method == "ssq" else rolling_causal_bands
for item in get_instruments_watchlist():
ticker = item["ticker"]
try:
hist = get_historical(ticker, period="1y", interval="1d")
if len(hist) < lookback + 32:
continue
values = [h["close"] for h in hist]
dates = [h["date"] for h in hist]
start_idx = max(lookback, len(values) - 60)
decomposed = decomposer(
values, dates,
start_idx=start_idx, lookback=lookback,
num_levels=num_levels, wavelet=wavelet, step=1,
)
if not decomposed["dates"]:
continue
price_at_signal = decomposed["original"][-1]
bands = decomposed["bands"]
for i, band in enumerate(bands):
series = band["series"]
if not series:
continue
slope = _compute_slope(series)
energy = band.get("energy")
kind: Optional[str] = None
direction: Optional[str] = None
params: Optional[Dict] = None
if extremum_cfg.get("enabled", True):
direction = detect_extremum_signal(series)
kind = "extremum" if direction else None
if not direction and level_cfg.get("enabled", True):
threshold_k = level_cfg.get("threshold_k", 2.0)
direction = detect_level_threshold_signal(series, threshold_k)
if direction:
kind, params = "level_threshold", {"threshold_k": threshold_k}
if not direction and trend_cfg.get("enabled"):
direction = detect_trend_flatten_signal(
series,
trend_cfg.get("trend_days", 10), trend_cfg.get("flatten_days", 5),
trend_cfg.get("trend_threshold_k", 1.0), trend_cfg.get("flatten_threshold_k", 0.3),
)
if direction:
kind = "trend_flatten"
params = {k: trend_cfg.get(k) for k in ("trend_days", "flatten_days", "trend_threshold_k", "flatten_threshold_k")}
if not direction and accel_cfg.get("enabled"):
accel_days = accel_cfg.get("accel_days", 3)
accel_threshold_k = accel_cfg.get("accel_threshold_k", 1.5)
direction = detect_acceleration_signal(series, accel_days, accel_threshold_k)
if direction:
kind, params = "acceleration", {"accel_days": accel_days, "accel_threshold_k": accel_threshold_k}
if not direction and cross_cfg.get("enabled"):
sec_idx = int(cross_cfg.get("secondary_band", 1))
if 0 <= sec_idx < len(bands) and sec_idx != i:
direction = detect_band_cross_signal(series, bands[sec_idx]["series"])
if direction:
kind, params = "band_cross", {"secondary_band": sec_idx}
if not direction and energy_cfg.get("enabled") and energy:
threshold_k = energy_cfg.get("threshold_k", 2.0)
direction = detect_level_threshold_signal(energy, threshold_k)
if direction:
kind, params = "energy_threshold", {"threshold_k": threshold_k}
results.append({
"ticker": ticker,
"band_label": band["label"],
"period_low_days": band.get("period_low_days"),
"period_high_days": band.get("period_high_days"),
"signal_kind": kind,
"direction": direction,
"price_at_signal": price_at_signal,
"slope": slope[-1],
"value": series[-1],
"energy": energy[-1] if energy else None,
"ridge_period_days": None,
"params_json": json.dumps(params) if params else None,
})
# Ridge — one row per ticker (ssq only), not per band
if method == "ssq" and decomposed.get("ridge_period_days"):
ridge_series = [v for v in decomposed["ridge_period_days"] if v is not None]
if ridge_series:
ridge_kind = None
ridge_direction = None
ridge_params = None
if ridge_cfg.get("enabled"):
threshold_k = ridge_cfg.get("threshold_k", 2.0)
ridge_direction = detect_level_threshold_signal(ridge_series, threshold_k)
if ridge_direction:
ridge_kind, ridge_params = "ridge_shift", {"threshold_k": threshold_k}
results.append({
"ticker": ticker,
"band_label": "ridge",
"period_low_days": None,
"period_high_days": None,
"signal_kind": ridge_kind,
"direction": ridge_direction,
"price_at_signal": price_at_signal,
"slope": None,
"value": None,
"energy": None,
"ridge_period_days": ridge_series[-1],
"params_json": json.dumps(ridge_params) if ridge_params else None,
})
except Exception:
continue # one bad ticker must not abort the whole scan
return results
def compute_and_save_wavelet_signals(run_id: str) -> List[Dict]:
from services.database import save_wavelet_signals
results = scan_watchlist_wavelet_signals()
save_wavelet_signals(run_id, results)
return results