From Stationarity to Deep Learning

A Comparative Framework for Stock Price and Volatility Forecasting

Author: Vishvas Ranjan
Affiliation: UM DAE Centre for Excellence in Basic Sciences, Mumbai, India
Supervisor: Prof. Amiya Bhowmick, Institute of Chemical Technology, Mumbai, India

---

Overview

This repository accompanies a detailed report (PDF) that builds a bridge from classical time‑series models (ARMA‑GARCH) to modern deep‑learning methods (LSTM) for forecasting:

• Closing prices of financial assets
• Volatility of those assets

We compare performance, highlight methodological synergies and divergences, and include full mathematical background, code snippets, and result discussions.

Note: Code quality is under improvement:— all codes attached in the pdf are executable. Separate Notebook will be provided asap!

---

Repository Contents

• PDF
  • Complete write‑up with:
    • Introduction to stationarity, ACF, ARMA processes, GARCH models
    • LSTM neural network theory
    • Comparative experiments (data preprocessing, model training, results)
    • Checkout Tables of contents for quick reference
• notebooks/ (coming soon)
  • Executable Jupyter notebooks for all code used in the report
  • Python scripts for data loading, model training & evaluation
  • All codes in the pdf are executable

---

Report Table of Contents (excerpt)

1. Time‑Series Model
   • Zero‑mean models, trends & seasonality
   • Stationary processes, ACF, MA(q) & AR(1)
   • ARMA processes & sample ACF properties
2. GARCH Process
   • Definition of GARCH(p,q)
3. LSTM Neural Network
4. Comparing Forecasting
   • LSTM model: preprocessing, training, visual analysis
   • ARMA‑GARCH method
   • Volatility prediction via ARMA‑GARCH & LSTM‑GARCH
   • Conclusions on synergy & divergence

---

Thanks for checking out!