Equilibrium-India-V1 (Space)

FastAPI backend for systemic risk prediction + simulation.

Equilibrium-India-V1

Equilibrium-India-V1 is a GRU-based machine learning model for systemic risk prediction in financial networks, designed for scenario analysis and stress testing of interconnected institutions.

The model predicts next-step systemic risk using network-theoretic and statistical features derived from an adjacency matrix representing financial interconnections.

---

🔍 Model Overview

• Model type: Gated Recurrent Unit (GRU)
• Framework: TensorFlow / Keras
• Task: Time-series regression
• Output: Predicted systemic risk score (continuous)
• Input window: Rolling lookback over historical system states

---

📥 Input Features

The model expects the following 6 derived features at each timestep:

Feature	Description
lambda_max	Largest eigenvalue of the financial network adjacency matrix
mean_risk	Mean node-level risk across institutions
max_risk	Maximum node-level risk (stress concentration)
std_risk	Risk dispersion across nodes
S_lag1	Systemic risk at previous timestep
S_lag5	Systemic risk 5 timesteps ago

These features are computed externally and passed to the model after scaling.

🧠 Training Details

• Loss function: Mean Squared Error (MSE)
• Optimizer: Adam
• Regularization: Dropout + L2 weight decay
• Early stopping: Enabled (best validation epoch restored)

The model was trained on historical market data combined with synthetic balance-sheet-style risk signals to approximate systemic exposure.

---

📊 Intended Use

This model is intended for:

• Systemic risk forecasting
• Financial network stress testing
• Scenario simulation (via downstream cascade models)
• Academic research and hackathon demonstrations

It is designed to be used as part of a larger pipeline, not as a standalone trading signal.

---

⚠️ Limitations

• Trained on a fixed institution universe
• Not suitable for arbitrary new stocks without retraining
• Synthetic risk proxies are approximations, not audited balance sheets
• Output is a relative risk indicator, not an absolute probability of failure

---

🚫 Not Intended For

• Investment advice
• Regulatory capital decisions
• Production financial risk management

---

🔗 Deployment

This model is consumed by a FastAPI-based inference service hosted as a Hugging Face Space, where it is used for real-time scenario simulations.

---

📜 License

MIT License

---

👤 Author

Vansh Momaya