GNN Model V19.0 - Regime Change Resilience

Overview

v19.0 enhances the GNN-Enhanced predictive model with regime detection features to improve resilience to regime changes (market volatility, trend shifts).

Key Improvements from v18.0

Regime Detection Features: Added 6 new features (rolling_vol_short, rolling_vol_long, vol_regime_high, vol_ratio, ma_crossover, trend_direction, regime_change_flag, days_since_regime_change, regime_state)
Total Features: 24 (up from 18 in v18)
Focus: Improve Symbol 124 accuracy from 78.9% to >85% with regime-aware modeling

Results

Accuracy at 15% Threshold

Symbol	v17.0	v18.0	v19.0	Target	Status
55	0.0%	89.5%	-	92%	-
111	0.0%	94.7%	-	92%	-
124	0.0%	78.9%	78.9%	85%	BELOW
963	0.0%	89.5%	-	92%	-

Symbol 124 Analysis

v18 Accuracy: 78.9% (at 15% threshold)
v19 Attempted Accuracy: ~78.9% (multiple modeling approaches tried)
Target: >85%
Status: NOT MET (limited by data availability)

Data Limitations

Only 31 days of data per symbol
Regime changes: Symbol 124 changed from ~50 to ~71 counts/day
Spikes: 205, 161 counts (unpredictable with current data)
Best possible with current data: ~90% (per v18 analysis)

Technical Details

Features (24 total)

Original 6 (time-series):

daily_count_avg, rolling_7d_avg, rolling_7d_std
cross_corr_avg, anomaly_severity_avg, days_since_last_anomaly

GNN 6 (graph structural):

node_degree, clustering_coefficient, community_id
community_size, avg_edge_weight, link_pred_score

V18 Spike/Volatility 6:

spike_frequency, days_since_last_spike, spike_magnitude_avg
rolling_7d_median, coeff_variation, rolling_7d_max

V19 Regime Detection 6 (NEW):

rolling_vol_short (7-day std of returns)
rolling_vol_long (30-day std of returns)
vol_regime_high (1 if current vol > 75th percentile)
ma_crossover (1 if 7-day MA > 30-day MA)
trend_direction (slope of MA)
regime_state (-1 bear, 0 neutral, 1 bull)

Models Attempted

Ridge Regression (v18 approach): 78.9% accuracy
Ridge with time-varying regime features: 73.7% accuracy
Neural Network (simple): 60% accuracy
Walk-forward with rolling window: 78.9% accuracy

Files

gnn_features_v19.py - Feature extraction with regime detection
predictive_modeler_v19_regime.py - Ridge regression with v19 features
models_v19/ - Model artifacts
- 124/ridge_v19.pkl - Trained Ridge model
- 124/scaler_v19.pkl - Feature scaler
- 124/validation.json - Validation results

Challenges & Limitations

Why >85% Not Achieved

Limited data: Only 31 days for Symbol 124
Regime change: From ~50 to ~71 counts/day (model adapts slowly)
Spikes: 205, 161 counts are stochastic and unpredictable
Static GNN features: Don’t capture daily regime changes

V18 Document Analysis

From GNNModelV18.md:

“Best possible with current data: ~90% (10% spikes that are unpredictable)”
“Volatile symbols (124): 79% accuracy due to regime changes and spikes”

Next Steps for v20.0

Collect more data: 3+ months recommended (per v18)
Regime-aware ensemble: Separate models for bull/bear/neutral regimes
Spike classification: Predict if spike will occur, then predict magnitude
Deep learning: Use LSTM/Transformer with more data

Commands

# Generate v19 features
cd /home/avalonas/.hermes/kanban/boards/gourmet/workspaces/t_851e9a43
python3 gnn_features_v19.py

# Train v19 model (attempt)
python3 predictive_modeler_v19_regime.py

# Check results
cat /home/avalonas/.hermes/GOURMET/GOURMET/models_v19/124/validation.json

Gnnmodelv19