/Predicting_stock_movement

Primary LanguageJupyter Notebook

Predicting_stock_movement

In this project, the historical APPLE stock daily Close price data is used to build various machine learning models to predict the stock price movement. The historical data includes all data since APPLE stock's launch in 1980-12-12 until 2022-02-04. The models built include

  • baseline model using mean value

  • smoothing model (single, double and triple exponential smoothing models)

  • SARIMA models of various orders using SARIMAX python library

  • SARIMA models of various orders using pmdarima python library

  • FB Prophet model

  • RNN model

  • LSTM model.

The models are trained using the historical data since its launch in 1980-12-12 until 2022-01-21, and the last ten day's data (from 2022-01-24 to 2022-02-04) were hold out for out-of-sample testing. Then the models' performances are compared using the metric MAPE (mean absolute percentage error).

The best model we found is a deep learning RNN model which outperforms all other models in making out-of-sample forecasts in the metric of MAPE. The RNN model we found slightly outperforms the LSTM model we built.

A double exponential smoothing model and a SARIMA model provide slightly worse performance but is much inexpensive to train.

A FB Prophet model and a SARIMA model built by the python library pmdarima perform significantly worse, which is somewhat surprising.

Meanwhile we note that there is a significant amount of randomness in training a deep learning model (either RNN or LSTM), even with the same set of hyper-parameters. Of course, training a deep learning model is much more expensive in terms of computing time.