Model	RMSE ↓	MAE	R²	Improvement
BiLSTM (this model)	6.38	3.82	0.556	Baseline
LSTM+Attention	9.19	5.37	0.264	-30.6%
Linear Regression	9.89	5.85	0.176	-35.5%
SARIMA	11.52	6.78	-0.090	-44.6%

Key Achievement: 37% improvement over best classical baseline (Linear Regression)

Model Architecture

Type: Bidirectional LSTM (2 layers)
Input Shape: (30 weeks, 13 features)
Output: Single value (next week's attack count)
Parameters: ~36,673
Framework: TensorFlow 2.13.0

Architecture Details:

    Input(30, 13) 
         ↓

Bidirectional LSTM(64) + Dropout(0.2) ↓ Bidirectional LSTM(32) + Dropout(0.2) ↓ Dense(32, ReLU) + Dropout(0.2) ↓ Dense(1, Linear)

Training Data

Dataset: Global Terrorism Database (START consortium)
Time Period: 1970-2016 (46 years)
Resolution: Weekly aggregation (2,400 weeks)
Train/Val/Test: 70%/15%/15% (chronological split)

Features (13 total):

Lag Features (1): 52-week lag
Rolling Statistics (4): 4-week & 12-week mean/std
Temporal Encoding (4): Year, week, month, quarter
Casualty Features (3): Killed, wounded, total
Geographic (1): Region encoding

Usage

python

import tensorflow as tf
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download(
    repo_id="Davidavid4/bilstm-terrorism-forecasting-gtd",
    filename="bidirectional_lstm_best.h5"
)

# Load model
model = tf.keras.models.load_model(model_path)

# Prepare input: shape (batch_size, 30, 13)
# - 30 weeks of historical data
# - 13 features per week

# Make predictions
predictions = model.predict(X_test)