Forecasting Trend Of S&P 500 ETF Index Using Recurrent Neural Networks(RNN)

This paper follows the in-sample forecasting criterion in applying artificial neural networks to SPY historical data,an index which tracks the S&P 500.The experiment has found that there's a general performance improvement in forecasting future prices using artificial neural networks over traditional statistical models,with both short term intervals(six months)and long-term intervals(5 years).By using broader predictors(high,low,open,and closing prices),the Neural Network was able to correctly predict the approximate values of future closing prices,which is a significant improvement from older forecasting methods of using Box-Jenkins method(ARIMA model).However,ARIMA model,used in time series forecasts,seem to perform better than simple machine learning(ML)linear regression algorithms,a linear approach that have not taken into the factor of time.This paper also proves that artificial neural networks are more efficient in general because a lack of assumption constraints;however,individual researchers do have to be careful in using such technology because of overfitting,one of the greater dangers of deep learning if applied incorrectly.This paper will also avoid using complex technical indicators approach in neural networks often producing mixed results based on the indicators selected.The primary goal of this research is not to prove that statistical modeling is obsolete in stock markets but rather to show the change in paradigm brought by newer technology on traditionally accepted views of predicting the stock market.This paper shows the differences of approach in-sample forecasting stock market data with traditional means(ARIMA model)and new technology-based means,machine linear modeling and artificial neural networks.While the performance tests of different methods are done using the same data,this paper has found that machine learning methods' performance is heavily based on two important variables: the algorithm selected based on the econometrics forecasting criterion and the amount of data available.Finally,to test the true difference between econometrics models and artificial neural networks this research will attempt to portray their differences absent any complex technical indicators,which often produces mixed results.