Time Series Forecasts using the Neural Networks

This diploma thesis is devoted to the analysis of the Long Short-Term Memory neural networks performance in prediction of randomly selected financial time series from Yahoo!Finance. Firstly,we determined five most precise LSTM models used for predictions of 150 financial time series of stocks from various industries. Then we calculated the statistical measures related to the time series predictability - the Hurst Coefficient, Metric Entropy and the largest Lyapunov Exponent. By estimation of the simple regression lines and the correlation coefficients we try to determine possible relationship between the quality of prediction represented by the average RMSE and each of the statistics related to the time series predictability with the consequent comparison with the theoretical assumptions.