Transformer-VAE-Stock-Predictor

A deep generative model combining Transformer-based Variational Autoencoders (VAE) and Normalizing Flows to generate and forecast financial time series data, conditioned on market volatility indicators.

---

Overview

Financial markets exhibit non-linear patterns, heavy-tailed distributions, and volatility clustering — making classical methods like ARIMA or GARCH insufficient for accurate modeling.

This project proposes a Transformer-VAE architecture with RealNVP normalizing flows, conditioned on the VIX volatility index, to generate synthetic financial data and learn latent financial dynamics.

---

Installation

Clone this repository and set up a Python environment:

git clone https://github.com/muralikarteek7/Transformer-VAE-Stock-Predictor.git
cd Transformer-VAE-Stock-Predictor
conda create -n transvae python=3.8
conda activate transvae

Dependencies

Install the following dependencies via pip or conda:

• torch
• transformers
• numpy
• pandas
• matplotlib
• yfinance
• scikit-learn

---

Dataset

Sources:

• Historical stock data from Yahoo Finance via yfinance
• Volatility Index (VIX) data

We used:

• 3 stocks: e.g., AAPL, MSFT, GOOGL
• Features: [Close, Volume]
• Time range: Multiple years (daily frequency)

Format:

Each sample is a sequence of shape (1000, 6):

• 1000 time steps
• 6 dimensions (3 stocks × 2 features each)
• VIX used as a conditional input

---

Model Architecture

• Loss: ELBO (Reconstruction + KL) + Flow log-det Jacobian

---

Evaluation Metrics

• Cosine Similarity: Between real and generated log returns
• Volatility clustering check
• Time-series overlay of real vs. synthetic returns

---

📊 Baseline Comparison

Model	Wasserstein Distance	Cosine Similarity
Transformer-VAE-Flow	0.0629	0.9072
TimeGrad	0.3246	0.9905
QuantGAN	0.0159	0.9148
DiM	0.2018	—
GARCH	0.1499	0.9919

---

Use Cases

• Synthetic data generation for backtesting
• Risk modeling and stress testing
• Portfolio simulation
• Improving data diversity for trading models

---

References

1. QuantGAN: Deep Generation of Financial Time Series
2. Bollerslev, T. “Generalized Autoregressive Conditional Heteroskedasticity.” Journal of Econometrics, 1986.
3. TimeGrad: Diffusion-based Forecasting