This paper aims to analyze if time series frequency is related to better forecast performance. We analyze two time series from G7 economies, NASDAQ and DAX, for 5 minutes, and daily frequency. The implemented algorithms are deep learning recurrent neural networks, particularly some variations of Long Short-Term Memory (LSTM) architectures (LSTM, BiLSTM). Random search hyperparameter tuning was used to obtain the model architecture that minimize the loss function. We obtained better results for the 5-minute intraday frequency for both time series, and the forecast improved by 1%.

--