DDPM-GARCH

DDPM-based financial return simulation framework for volatility forecasting, Value-at-Risk (VaR) estimation, and Monte Carlo comparison using GARCH(1,1).

This repository accompanies the research paper:

DDPM as a Data-Driven Return Simulator: Benchmarking Financial Risk Analysis Against GARCH and Monte Carlo Baselines

---

Features

• DDPM implementation for financial return generation
• 1D U-Net architecture for time-series diffusion
• Pretrained models for multiple window lengths
• Forward and reverse diffusion process visualization
• GARCH(1,1) volatility modeling
• Monte Carlo simulation baseline
• Value-at-Risk (VaR) estimation
• Volatility forecasting evaluation
• FastAPI web application
• Simple HTML showcase interface

---

Pretrained Models

The repository contains three pretrained models:

Model	Window Length
model_128.pth	128 returns
model_256.pth	256 returns
model_512.pth	512 returns

Each model generates fixed-length synthetic return windows using the reverse diffusion process.

---

Project Structure

.
├── app/
│   ├── main.py
│   └── ...
│
├── static/
│   ├── index.html
│   └── ...
│
├── models/
│   ├── model_128.pth
│   ├── model_256.pth
│   └── model_512.pth
│
├── diffusion/
│   ├── forward_process.py
│   ├── reverse_process.py
│   └── ...
│
├── garch/
│   └── ...
│
├── data/
│   └── ...
│
└── README.md

---

Installation

Clone the repository:

git clone https://github.com/HenryGunadi/diffusion-garch.git
cd diffusion-garch

Create a virtual environment:

python -m venv .venv

Activate the environment:

Windows

.venv\Scripts\activate

Linux / macOS

source .venv/bin/activate

Install dependencies:

pip install -r requirements.txt

---

Running the Web Application

From the project root:

uvicorn app.main:app --reload

Open:

http://127.0.0.1:8000

The web interface is served through the FastAPI backend and provides a simple demonstration of synthetic return generation.

---

Diffusion Workflow

Forward Process

The forward process gradually adds Gaussian noise to return windows until they become nearly pure noise.

Historical Returns
        │
        ▼
Forward Diffusion
        │
        ▼
Gaussian Noise

Reverse Process

The reverse process starts from Gaussian noise and iteratively reconstructs synthetic return windows.

Gaussian Noise
        │
        ▼
Reverse Diffusion
        │
        ▼
Synthetic Returns

---

Monte Carlo Baseline

The repository also includes a traditional Monte Carlo benchmark.

Workflow:

Historical Returns
        │
        ▼
Fit GARCH(1,1)
        │
        ▼
Monte Carlo Simulation
        │
        ▼
Synthetic Returns

This baseline is used for comparison against DDPM-generated scenarios.

---

Example Workflow

Generate Synthetic Returns

1. Load a pretrained model.

model_512.pth

2. Generate synthetic return windows using reverse diffusion.

3. Fit a GARCH(1,1) model.

4. Perform:

• Volatility forecasting
• VaR estimation
• Risk analysis
• Monte Carlo comparison

---

Models

Model	Recommended Use
model_128.pth	Short horizon experiments
model_256.pth	Medium horizon experiments
model_512.pth	Long horizon experiments

---

Showcase

The repository includes:

• main.py — FastAPI backend
• index.html — lightweight web showcase

The showcase demonstrates:

• Loading pretrained models
• Running reverse diffusion sampling
• Generating synthetic return windows
• Comparing DDPM outputs against GARCH-based simulations

---

Research

The full paper is available in this repository.

Topics covered include:

• Stylized facts retention
• GARCH diagnostics
• Volatility persistence
• Volatility forecasting
• QLIKE evaluation
• Value-at-Risk estimation
• Monte Carlo benchmarking

---

Disclaimer

This project is intended for research and educational purposes only.

It should not be used as financial advice or investment recommendations.