Prediction Of The Daily Closing Value Of The S&P 500 Based On Deep Learning

Stock market prediction has long been a focus of investors,traders,and economists.Economic forecasting allows economists to make suggestions for government officials and business owners,yet current models are mainly based on inflation,interest rates,industrial production,consumer confidence,worker productivity,retail sales and unemployment rates.This paper focuses on another important aspect of economic forecasting: stock market trends.In our research,we predict the value of the S&P 500 market index 30 days into the future using stateof-the-art deep learning algorithms.This research is not intended to create stock market trading algorithms.Rather,this research is intended to forecast economic trends in the S&P 500 Index.In our research we use three of these deep learning algorithms.They are compared on equal footing and use the same datasets and parameters wherever possible.Our first algorithm is a long short-term memory(LSTM)algorithm.The second is a modified gated recurrent unit(GRU),and our third model is a bidirectional GRU(Bi-GRU).In our research,we show the enhanced accuracy of our Bi-GRU algorithmic model.As shown in the paper,the models gain enhanced predictive power as these features evolve from LSTM to GRU to Bi-GRU,each step of the way the enhancements in our programming lead to better predictability and accuracy in our models.Most research in machine learning applications of stock market forecasting use LSTM algorithms to make predictions.We take the latest Tensor Flow algorithmic structures and apply them to historical data for the S&P 500 index and attempt to predict the future closing value of the index 30 days into the future using publicly available datasets from Yahoo finance.Our findings are promising.We find that gated recurrent units outperform LSTM algorithms in S&P 500 forecasting.We also show that it is possible to forecast the future of the S&P 500 index with reasonable accuracy,and we control for overfitting by exposing our models to different datasets and achieve similar results.This shows that our results are accurate.Our main finding is that bidirectional GRUs outperform unidirectional GRUs,and we postulate that this is due to the increased connectivity of our neural networks in Bi-GRU algorithmic models.