Time Series Forecasting with Hybrid Models

This project involves time series forecasting for stock prices using a hybrid approach combining ARIMA and LSTM models.

Overview

The project utilizes historical stock data from 'GOOG.csv' and employs the following steps:

1. Data Preprocessing

   • Visualize the original closing prices.
   • Make the data stationary by taking the first difference.
   • Analyze Fourier Transform to check for stationarity.

2. Autocorrelation Analysis

   • Use ACF and PACF plots to determine model parameters (p, d, q) for ARIMA.

3. ARIMA Modeling

   • Fit an ARIMA model on the training data using identified parameters.
   • Forecast stock prices on the test data.
   • Plot the original test data and forecasted values.

4. Error Analysis

   • Calculate prediction errors between actual and forecasted values.
   • Visualize prediction errors.

5. LSTM Modeling for Error Prediction

   • Train an LSTM model on the errors from the ARIMA model.
   • Create sequences of errors for training the LSTM model.

6. LSTM Model Training

   • Build and train an LSTM model to predict errors.
   • Evaluate the performance of the LSTM model.

7. Overall Prediction

   • Combine ARIMA and LSTM predictions to get the final forecast.

Requirements

• Python 3.x
• Libraries: TensorFlow, Statsmodels, Pandas, NumPy, Matplotlib, Scikit-learn
• One might require to download the tensorflow library as it is not inbuilt in base kernel

Results

• The project aims to predict stock prices using a hybrid ARIMA-LSTM model.
• The performance of the model can be evaluated through visualizations and metrics.

Acknowledgments

• The project is a part of the WIDS workshop (Winter in Data Science)
• Yahoo Finance for their historical data regarding stock prices