The prediction of complex stock dynamics is a challenging task. Prices in developed stock markets are among the most complex financial data available to model since their fluctuations are irregular, nonlinear and change dynamically with high uncertainty. This paper discusses and analyses different models for stock market price prediction, namely S&P 500 from 1964 until 2020, exploiting recurrent and long short-term memory (LSTM) deep learning algorithms. In recent years, a variety of research fields, including finance, have begun to place great emphasis on machine learning techniques because they exhibit improved performance in such stochastic simulations. Forecasting share performance becomes a more challenging issue due to the enormous amount of valuable trading data stored in the stock databases. Existing classic forecasting methods are insufficient to analyse the share performance accurately. The neural network models, and in particular, the LSTM algorithm, has been increasingly used by researchers for prediction and analysis of stock market data, with an important role in today’s economy. Stock market index prediction is an example, and by modifying the definitions of input variables and output of each module, the same framework as the proposed method can be applied to various deep learning-based financial problems, including financial market movement classification, volatility prediction, portfolio optimisation, and option pricing. The majority of empirical work deal with daily data since they could collect a bigger dataset to analyse. However, this issue of daily data can be largely dependent on deterministic variables like day-of-the-week, week-of-the-year, month of-the-year, week-of-the-month, long-weekends. Furthermore, there may be changes in daily patterns over time and different volatilities (uncertainties/variability) for different days of the week. For this reason, we decide in this work, to introduce an attempt for prediction of stock market trend using different frequencies. Three models are built: the first one address daily prediction, the second one is for weekly data and finally the third model concern monthly prediction. Different hyperparameter tunning and cross-validation are used for each network architecture. We have found that the LSTM configuration works the best out of all the combinations we have tried (for our dataset) for weekly data, one hidden layer and dropout regularization. LSTMs are very powerful in sequence prediction problems because they can store past information, which is crucial in predicting its future price and/or trend. We use Python software, where Keras, TensorFlow and Pandas are the packages with the leading role.

--