The paper aims to assess if the frequency of time series is associated with increased forecast accuracy. We examine two different time series from the G7 countries, the NASDAQ100 and the DAX, for a period of five minutes, as well as daily frequency. The employed algorithms are deep learning recurrent neural networks that are particularly suited for a variety of variations of Long Short-Term Memory (LSTM) structures (LSTM, BiLSTM). A random search over the hyperparameters was employed to determine the architecture that minimizes the loss function. We had a better outcome for the 5-minute daily frequency for both datasets, the forecast increased by 1%.

--

Recurrent neural models,Stock markets,Forecasting,Time series