Spaces:
Running
Running
File size: 7,329 Bytes
e57e9d1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 | """
Advanced Sentiment Feature Engineering.
Computes second-order derivatives of sentiment signals that capture
market-moving dynamics the raw sentiment_index cannot:
- Sentiment Momentum (SMA/EMA rate-of-change)
- Sentiment Surprise (Z-score anomaly detection)
- Volume-Weighted Sentiment
- Event-Type Intensity (supply_disruption counts, etc.)
"""
from __future__ import annotations
import logging
from typing import Optional, Sequence
import numpy as np
import pandas as pd
from deep_learning.config import SentimentFeatureConfig, get_tft_config
logger = logging.getLogger(__name__)
def compute_sentiment_momentum(
sentiment_index: pd.Series,
windows: Sequence[int] = (5, 10, 30),
) -> pd.DataFrame:
"""
Sentiment momentum: how fast and in which direction market mood is shifting.
For each window *w* the feature set contains:
- sent_momentum_{w}d : current index minus SMA(w)
- sent_ema_{w}d : EMA(w) of the index (smoothed trend)
- sent_roc_{w}d : rate-of-change over w days
"""
features = pd.DataFrame(index=sentiment_index.index)
for w in windows:
sma = sentiment_index.rolling(window=w, min_periods=1).mean()
features[f"sent_momentum_{w}d"] = sentiment_index - sma
features[f"sent_ema_{w}d"] = sentiment_index.ewm(span=w, adjust=False).mean()
features[f"sent_roc_{w}d"] = sentiment_index.diff(w)
return features
def compute_sentiment_surprise(
sentiment_index: pd.Series,
lookback: int = 30,
threshold: float = 2.0,
) -> pd.DataFrame:
"""
Sentiment Surprise: Z-score of today's sentiment relative to recent history.
When |z| >= threshold the market received an "unexpected" signal that
historically triggers outsized price moves.
Features:
- sent_surprise_z : rolling Z-score
- sent_surprise_flag : binary flag (|z| >= threshold)
- sent_surprise_signed : z * sign (directional surprise magnitude)
"""
roll_mean = sentiment_index.rolling(window=lookback, min_periods=5).mean()
roll_std = sentiment_index.rolling(window=lookback, min_periods=5).std()
roll_std = roll_std.replace(0, np.nan)
z_score = (sentiment_index - roll_mean) / roll_std
features = pd.DataFrame(index=sentiment_index.index)
features["sent_surprise_z"] = z_score
features["sent_surprise_flag"] = (z_score.abs() >= threshold).astype(np.float32)
features["sent_surprise_signed"] = z_score * np.sign(sentiment_index)
return features
def compute_volume_weighted_sentiment(
sentiment_index: pd.Series,
news_count: pd.Series,
) -> pd.DataFrame:
"""
Weight sentiment by news volume: high-volume days carry stronger signal.
Features:
- sent_vol_weighted : sentiment * log(1 + news_count)
- sent_vol_zscore : Z-score of volume-weighted series (30-day)
- news_count_zscore : Z-score of news volume itself
"""
log_count = np.log1p(news_count.fillna(0))
vol_weighted = sentiment_index * log_count
vol_roll_mean = vol_weighted.rolling(30, min_periods=5).mean()
vol_roll_std = vol_weighted.rolling(30, min_periods=5).std().replace(0, np.nan)
nc_roll_mean = news_count.rolling(30, min_periods=5).mean()
nc_roll_std = news_count.rolling(30, min_periods=5).std().replace(0, np.nan)
features = pd.DataFrame(index=sentiment_index.index)
features["sent_vol_weighted"] = vol_weighted
features["sent_vol_zscore"] = (vol_weighted - vol_roll_mean) / vol_roll_std
features["news_count_zscore"] = (news_count - nc_roll_mean) / nc_roll_std
return features
def compute_event_type_intensity(
event_counts: pd.DataFrame,
event_types: Optional[Sequence[str]] = None,
) -> pd.DataFrame:
"""
Daily counts per event-type category from NewsSentimentV2.
*event_counts* is expected to have date index and one column per event_type
with daily occurrence counts.
Features (per event type):
- evt_{type}_count : raw daily count
- evt_{type}_ma5 : 5-day moving average
- evt_{type}_spike : flag when count > 2 * MA5
"""
if event_types is None:
cfg = get_tft_config()
event_types = list(cfg.sentiment.event_types)
features = pd.DataFrame(index=event_counts.index)
for etype in event_types:
col = etype if etype in event_counts.columns else None
if col is None:
features[f"evt_{etype}_count"] = 0.0
features[f"evt_{etype}_ma5"] = 0.0
features[f"evt_{etype}_spike"] = 0.0
continue
counts = event_counts[col].fillna(0).astype(float)
ma5 = counts.rolling(5, min_periods=1).mean()
features[f"evt_{etype}_count"] = counts
features[f"evt_{etype}_ma5"] = ma5
features[f"evt_{etype}_spike"] = (counts > 2.0 * ma5.clip(lower=0.5)).astype(np.float32)
return features
def build_event_counts_from_db(
session,
start_date,
end_date,
) -> pd.DataFrame:
"""
Query NewsSentimentV2 to build a (date x event_type) count matrix.
"""
from sqlalchemy import func as sa_func
from app.models import NewsSentimentV2, NewsProcessed, NewsRaw
rows = (
session.query(
sa_func.date(NewsRaw.published_at).label("date"),
NewsSentimentV2.event_type,
sa_func.count(NewsSentimentV2.id).label("cnt"),
)
.join(NewsProcessed, NewsSentimentV2.news_processed_id == NewsProcessed.id)
.join(NewsRaw, NewsProcessed.raw_id == NewsRaw.id)
.filter(NewsRaw.published_at >= start_date, NewsRaw.published_at <= end_date)
.group_by(sa_func.date(NewsRaw.published_at), NewsSentimentV2.event_type)
.all()
)
if not rows:
return pd.DataFrame()
records = [{"date": r.date, "event_type": r.event_type, "count": r.cnt} for r in rows]
df = pd.DataFrame(records)
pivot = df.pivot_table(index="date", columns="event_type", values="count", fill_value=0)
pivot.index = pd.to_datetime(pivot.index)
return pivot
# ---------------------------------------------------------------------------
# Unified builder
# ---------------------------------------------------------------------------
def build_all_sentiment_features(
daily_sentiment: pd.DataFrame,
event_counts: Optional[pd.DataFrame] = None,
cfg: Optional[SentimentFeatureConfig] = None,
) -> pd.DataFrame:
"""
Build the complete sentiment feature set from daily_sentiment DataFrame.
Expected columns in *daily_sentiment*:
- sentiment_index (float)
- news_count (int)
"""
if cfg is None:
cfg = get_tft_config().sentiment
si = daily_sentiment["sentiment_index"]
nc = daily_sentiment["news_count"]
parts: list[pd.DataFrame] = [
compute_sentiment_momentum(si, windows=cfg.momentum_windows),
compute_sentiment_surprise(si, lookback=cfg.surprise_lookback, threshold=cfg.surprise_threshold),
compute_volume_weighted_sentiment(si, nc),
]
if event_counts is not None and not event_counts.empty:
evt_aligned = event_counts.reindex(daily_sentiment.index).fillna(0)
parts.append(compute_event_type_intensity(evt_aligned, event_types=cfg.event_types))
combined = pd.concat(parts, axis=1)
return combined
|