A Case Study in Session-Aware Bullish Hammer Detection
How session-aware volume feature engineering solved the critical problem of false volume spikes and dramatically improved ML model performance for ES futures pattern recognition. This breakthrough eliminated 228 false signals (21% of patterns) and transformed volume features from 0% to 3.5% importance.
False Volume Spikes from Session Mixing
Initial ML model completely ignored volume features:
Volume features were getting overshadowed by price-based features, indicating a fundamental problem with volume calculation methodology.
User Discovery: "The overnight volume is generally much lower than the cash session - so mixing overnight volumes and cash session will give false results."
When market transitions from overnight (~700 volume) to cash session (~15,000 volume), traditional ratios create false 4.6x spikes that aren't actually anomalous.
Breakthrough Engineering Approach
Separate trading sessions by time boundaries:
Different volume thresholds for each session:
Accounts for different baseline volume levels across sessions.
24-Hour Session Classification for Volume Analysis
def classify_session(hour, minute):
time_minutes = hour * 60 + minute
if 8 * 60 + 30 <= time_minutes <= 15 * 60: # 8:30 AM - 3:00 PM
return 'CASH'
elif time_minutes >= 17 * 60 or time_minutes < 8 * 60 + 30: # 5:00 PM - 8:30 AM
return 'OVERNIGHT'
else:
return 'AFTER_HOURS' # 3:00 PM - 5:00 PM# Process each session separately
for session in ['CASH', 'OVERNIGHT', 'AFTER_HOURS']:
session_mask = df['session'] == session
session_data = df[session_mask]
# Volume averages WITHIN SESSION ONLY
df.loc[session_mask, 'volume_ratio_10_session'] = (
session_data['Volume'] / session_data['Volume'].rolling(10).mean()
)
df.loc[session_mask, 'volume_ratio_20_session'] = (
session_data['Volume'] / session_data['Volume'].rolling(20).mean()
)# Different thresholds for different sessions
if session == 'CASH':
# Higher thresholds for cash session (more volume normally)
moderate, high, very_high, extreme = 1.5, 2.0, 3.0, 4.0
else:
# Lower thresholds for overnight (less volume normally)
moderate, high, very_high, extreme = 1.3, 1.8, 2.5, 3.5
session_data['is_moderate_volume_session'] = (session_data['volume_ratio_20_session'] >= moderate).astype(int)
session_data['is_high_volume_session'] = (session_data['volume_ratio_20_session'] >= high).astype(int)Session-Aware Model Success
| Feature | Traditional Model | Session-Aware Model | Improvement |
|---|---|---|---|
| volume_quality_score_session | 0% | 3.5% (#10 overall) | +3.5% |
| volume_ratio_20_session | 6.1% | 2.8% (#12 overall) | Stable ranking |
| is_high_volume_session | 0% | 2.1% (#15 overall) | +2.1% |
| False Volume Spikes | 228 patterns | 0 patterns | 100% eliminated |
All 12 high-confidence trades occurred during OVERNIGHT session, validating that the session-aware model correctly learned that overnight hammer patterns with volume spikes are more reliable than cash session patterns.
The session-aware model learned market microstructure: overnight hammer patterns are more reliable when they have true volume spikes, while cash session patterns are noisier despite higher absolute volume.
Complete Session-Aware Template
def engineer_session_aware_volume_features(df):
"""Session-aware volume feature engineering template"""
# 1. Classify trading sessions
df = add_session_classification(df)
# 2. Process each session separately
for session in ['CASH', 'OVERNIGHT', 'AFTER_HOURS']:
session_mask = df['session'] == session
session_data = df[session_mask].copy()
if len(session_data) == 0:
continue
# Session-specific volume averages (NO CROSS-SESSION CONTAMINATION)
session_data['volume_sma_20_session'] = session_data['Volume'].rolling(20, min_periods=1).mean()
session_data['volume_ratio_20_session'] = session_data['Volume'] / session_data['volume_sma_20_session']
# Session-specific percentiles (within session historical data only)
session_data['volume_percentile_session'] = session_data['Volume'].rolling(100, min_periods=5).rank(pct=True)
# Session-specific thresholds (different for each session type)
if session == 'CASH':
high_threshold, very_high_threshold = 2.0, 3.0 # Higher for cash
else:
high_threshold, very_high_threshold = 1.8, 2.5 # Lower for overnight
session_data['is_high_volume_session'] = (session_data['volume_ratio_20_session'] >= high_threshold).astype(int)
session_data['is_very_high_volume_session'] = (session_data['volume_ratio_20_session'] >= very_high_threshold).astype(int)
session_data['is_top_volume_decile_session'] = (session_data['volume_percentile_session'] >= 0.90).astype(int)
# Update original dataframe
for col in ['volume_ratio_20_session', 'is_high_volume_session', 'is_very_high_volume_session', 'is_top_volume_decile_session']:
if col in session_data.columns:
df.loc[session_mask, col] = session_data[col]
# 3. Session-aware composite score
df['volume_quality_score_session'] = (
df['is_high_volume_session'] * 2 +
df['is_very_high_volume_session'] * 3 +
df['is_top_volume_decile_session'] * 2
)
return dfdef add_session_classification(df):
"""Classify ES futures trading sessions"""
df['hour'] = df['Date'].dt.hour
df['minute'] = df['Date'].dt.minute
def classify_session(hour, minute):
time_minutes = hour * 60 + minute
if 8 * 60 + 30 <= time_minutes <= 15 * 60:
return 'CASH' # 8:30 AM - 3:00 PM CT
elif time_minutes >= 17 * 60 or time_minutes < 8 * 60 + 30:
return 'OVERNIGHT' # 5:00 PM - 8:30 AM CT
else:
return 'AFTER_HOURS' # 3:00 PM - 5:00 PM CT
df['session'] = df.apply(lambda row: classify_session(row['hour'], row['minute']), axis=1)
return dfdef identify_high_volume_hammers(df):
# Volume requirement integrated into pattern detection
is_bullish_hammer = (
lower_wick >= 2 * body and # Classic hammer shape
close > open and # Bullish bias
body > 0 and # Meaningful body
upper_wick <= body and # Clean rejection
total_range > 1 and # Sufficient range
volume_ratio >= 1.5 # VOLUME SPIKE REQUIRED
)
return is_bullish_hammerFor Quantitative Developers
The biggest breakthrough wasn't traditional feature engineering, but respecting market microstructure. Mixing overnight and cash session data creates systematic false signals that undermine model performance.
User feedback identified the core issue: "mixing overnight and cash session will give false results". This domain insight led to the session-aware solution that eliminated 228 false volume spikes.
Volume analysis must account for:
All 12 high-confidence trades occurred during overnight session, revealing that:
Session-aware volume feature engineering solved the critical problem of false volume spikes and transformed bullish hammer detection into a market structure-aware system. The breakthrough wasn't traditional feature engineering, but respecting the fundamental difference between overnight and cash session trading dynamics.
This approach is applicable to any time-series ML system where market microstructure matters. Traditional feature engineering ignores session boundaries at the cost of model robustness.
Author: Bibhash Biswas
Date: October 2025
Market: E-mini S&P 500 Futures (ES)
Framework: Python, scikit-learn, pandas
Key Innovation: Session-Aware Volume Feature Engineering