A temporal-attribute attention neural network for mixed frequency data forecasting

One of the dilemmas faced by forecasting is that the data are collected at different frequencies in some practical applications. This paper treats mixed frequency data as a special kind of multi-source data, that is, data from each source is collected at a different sampling frequency. Based on this cognition, this paper draws on the idea of neural network processing multi-source data to proposes a temporal-attribute attention neural network (TAA-NN for short) model to study the raw mixed frequency data. The new method avoids the problems caused by frequency alignments, such as information loss and artificial assumption of data distribution. First, this paper proposes a new sliding window strategy for mixed frequency data to determine the amount of data input into the model from each source data. Then, a group of convolutional neural network (CNN) with the same number of filters is used to extract or expand temporal features from the hidden state for each source data (a frequency data), so as to realize the information fusion of mixed frequency data at the feature layer. In addition, a temporal-attribute attention mechanism is proposed to mine essential information from the fused feature matrix in temporal and attribute dimensions. Experiments on two simulations and real-world datasets demonstrate that TAA-NN outperforms the compared methods and provides a new solution to the mixed frequency data forecasting.