Overview

Developed an end-to-end time series forecasting system to predict adjusted closing prices for five diverse stocks (MARA, SOXL, TD_BANK, NVDA, MANULIFE). This project involved data ingestion, exploratory data analysis (EDA), time series decomposition, and the implementation and comparative evaluation of multiple forecasting models: ARIMA, Exponential Smoothing (ETS), Prophet, and LSTM (using TensorFlow/Keras). Key findings highlight the significant variability in model performance across different stocks and the critical importance of hyperparameter tuning and data preprocessing techniques like Winsorization. The project also involves building an interactive Streamlit web application allowing users to upload their own stock data, select models, and visualize forecasts.

Problem Statement & Goal

The primary goal was to forecast the future adjusted closing prices for stocks representing different market sectors and volatility profiles. A secondary objective was to compare the effectiveness and suitability of various statistical, machine learning, and deep learning time series models (ARIMA, ETS, Prophet, LSTM) for this financial forecasting task under specific tuning strategies and data conditions.

Key Features & Functionality

• Data ingestion from multi-sheet Excel files.
• Automated Exploratory Data Analysis (Time series plots, histograms, outlier detection).
• Time Series Decomposition (Additive & Multiplicative) visualization.
• Implementation of four forecasting models: ARIMA (with auto-tuning), ETS (with grid search tuning), Prophet (default), LSTM (with stabilization).
• Training/testing split methodology for model evaluation.
• Calculation and display of standard forecast accuracy metrics (RMSE, MAE, MAPE, MSE).
• Generation of short-term (test set) and longer-term (future horizon) forecast visualizations.
• Interactive Streamlit Web Application for user-driven analysis and forecasting.

Technology Stack

• Python
• Pandas
• NumPy