ST311 Proposal
Objective
In finance, the concept of uncertainty is associated with volatility. Future volatility of S&P 500 is
chosen to be predicted with the time series model, Generalized AutoRegressive Conditional
Heteroskedasticity (GARCH) and Long Short-Term Memory (LSTM), which is a variation of
Recurrent Neural Network. Lastly, compare the forecast performance of GARCH model, LSTM
baseline model and LSTM extended model. Adam optimizer would be used for LSTM. The
chosen loss metric is Root Mean Squared Error (RMSE).
Source of dataset
From yfinance package, download the daily closing price of S&P 500 from 2018 to 2022. The
closing price is not stationary. Return is preferred as it is more stationary. When return is plot
against time, it shows no trend and tends to cluster. log-return would be computed with the
formula as follows:
Data Splitting
2 types of data splitting strategy would be explored to estimate the optimal hyperparameters
1. Time series cross validation
● Differently sized training data with equally sized validation data
2. Rolling time series cross validation
● Equally sized training and validation data