All Times ET
Traditional Time Series modeling has been proven very effective in the past. However, with the new data science world being expanding in the past decade, Artificial Intelligence or Deep Learning has become a very famous way of model fitting that yields great prediction accuracy. It is always a trade off between the prediction accuracy and the interpretability of the models fitted and when more flexible models as the deep learning models are implemented, we are giving up the interpretability of our model. This paper will compare the traditional time series model fitting vs. implementing Deep Learning models for forecasting financial time series data. This will be illustrated using stock prices of Tech companies in the US such as Apple Inc., Amazon (Amazon.com Inc.), Facebook(Meta Platforms Inc), Microsoft Corporation , Google(Alphabet Inc.) and Tesla Inc. The traditional model fitting will be done using models such as ARIMA,ARIMAX,GARCH, GJR-GARCH and VAR. For the comparison between the traditional time series modeling and deep learning modeling, Recurrent Neural network (RNN) and LSTM models will be used since time series data are sequential data. The deep learning models are very powerful in forecasting the next n time steps using the previous n time steps, especially when the time steps associated with the RNN models were apart from 1 minute intervals. Additionally, the deep learning models can be made into binary classifiers to predict if the closing price will increase (close up) or decrease(close down) than in the previous price time step. Usual Machine Learning models cannot be applied (except deep learning models) to Time Series data because Time series data contains correlation among their lagged observations/ the lagged series but in most of the ML models it is assumed that the observations are independently and identically distributed. This assumption will no longer be valid for time series data. Therefore, this very correlation among the lag observations/series are being used for time series model fitting. Forecasting stock prices is a very difficult and challenging task in the financial market because the trends of stock prices are non-linear and non-stationary time-series data. In financial time series data, there's clustering volatility presence in the data in addition to the correlation among the series itself, due to many world factors affecting the financial market of a country. Therefore, correlation among conditional variation of the returns will be used to model these volatility clustering in these stock prices by using GARCH and GJR-GARCH Models. This research will be extended to Forecast and compare stock market behavior with social media influence using sentiment analysis and time series analysis (using Twitter data, NewsAPIs and Stockwits APIs).