Hiruni2207/DataSynthis_ML_JobTask

📈 Stock Price Forecasting - DataSynthis ML Job Task

This repository contains implementations of time-series forecasting for stock prices using both traditional statistical models (ARIMA, Prophet) and deep learning (LSTM).
The project demonstrates model comparison, rolling-window evaluation, and deployment to Hugging Face Hub.

Project Overview

• Dataset: Daily stock price dataset (closing prices).
• Models Implemented:
  • ARIMA (AutoRegressive Integrated Moving Average)
  • Prophet (Additive Time Series Forecasting by Meta)
  • LSTM (Long Short-Term Memory Neural Network)
• Evaluation:
  • Rolling-window forecasts
  • Metrics: RMSE, MAPE
• Deployment:
  • Models and results shared on Hugging Face Hub.

Repository Contents

• lstm_model.h5 – Trained LSTM model
• scaler.pkl – Scaler used for preprocessing
• performance_summary.csv – Comparison of ARIMA, Prophet, and LSTM performance
• stock_forecasting_notebook.ipynb – Full notebook with preprocessing, training, evaluation, and plots
• upload_to_hf.py – Script for uploading to Hugging Face Hub

Quick start

1. Create and activate a python environment (recommended: conda or venv)

   python -m venv venv
   source venv/bin/activate   # Linux/macOS
   venv\Scripts\activate    # Windows
   pip install -r requirements.txt
   

2. Start Jupyter and open the notebook:

   jupyter notebook stock_forecasting_notebook.ipynb
   

3. The notebook contains cells to download real stock data via yfinance (if you have internet) or use the included sample_stock.csv for an offline demo.

Hugging Face deployment (notes)

• Use upload_to_hf.py to push saved model files to the HF repo DataSynthis_ML_JobTask after creating it on the Hugging Face website (or the script will create the repo for you if you provide a valid token).
• Create a HF token at https://huggingface.co/settings/tokens and set environment variable HF_TOKEN or pass --token to the script.

Results

The performance of the three models on stock price forecasting is summarized below:

Model	RMSE	MAPE (%)
ARIMA	3.3748	1.8973
Prophet	4.7650	3.1859
LSTM	2.0890	1.2516

Key Insights

• LSTM achieved the lowest RMSE and MAPE, showing the best accuracy.
• ARIMA performed reasonably well, but less effective with non-linear trends.
• Prophet captured trends and seasonality but had higher errors.
• Overall, LSTM is the most reliable model for this task.