Deep Learning For Time Series Forecasting: A Case Of Pharmaceutical Sales

Abstract

The emerging necessity for mathematical prediction of future behavior in numerous scientific and practical domains, coupled with the growing abundance of historical data, entails the conception of reliable and efficient techniques capable of capturing potential relationship between past and future from multivariate observations. However, there is a variety of methods utilized in predicting single variable data and cannot extract adequate data sequences between multiple variables. To address this limitation, a novel graph neural network model is introduced for multivariate time series forecasting specifically tailored to pharmaceutical sales forecasting. Each network node interconnected through hidden dependencies is used to exemplify a specific variable in the heterogeneous time series. Our proposed model utilizes intra-series temporal correlations extraction and attention mechanisms to capture the inter-series correlations in the graph. Wavelet transform decomposition is used to obtain time-related patterns from the multivariate time series at different frequencies, and simultaneously forming node features in the graph learning module. Multi-head attention is applied to integrate associations between nodes and enhance the relational dependencies in the graph structure. Additionally, graph convolutional neural networks update node embeddings for capturing abundant relationships. Finally, a temporal convolutional network is employed to establish correlation among temporal scales to achieve multivariate time series prediction. The effectiveness of our proposed model is evaluated using real-world data from medical domains, showing its superiority over existing deep learning methods. The proposed graph neural network model has the potential to advance multivariate time series forecasting in various applications.