Machine learning for time series analysis and forecasting
Ming Luo · 2023
We are immersed in a world with all types of data. Time series data are prevalent and essential in decision-making. Time series data have intrinsic temporal order and are thus immutable over time. The autocorrelation among time series data over time index makes them unique to deal with. Moreover, latent explanatory variables behind the time series make it challenging to handle. In this thesis, the author applies machine learning techniques to analyze time series data for classification, clustering, and forecasting. First, a new distance measure, value-added, is proposed in time series classification and clustering. Further, the author develops a novel framework in which decisions such as the number of clusters and prediction based on value-added are made using different techniques. Numerical real-world data studies demonstrate the value-added framework in time series classification and clustering. Forecasting in scale is a particular issue in business forecasting. Frequent forecasting with a hierarchy of time series data differs entirely from that with a single univariate time series. The author first reviews standard hierarchical time series forecasting methods. Then a new approach reconciliation with neural network (R-NN) is proposed for hierarchical forecasting, considering the non-linear relationship among time series for forecasting. Conventional techniques have yet to incorporate the relationship between series while making individual predictions and tend to lose information. In addition, computational costs can be exceptionally high due to model perplexity when using methods such as optimal reconciliation. The R-NN is straightforward to implement and fast to train without losing information. Numerical studies are shown, and accuracy improvements are observed. The methods and procedures developed in this thesis can be applied in various business settings. For example, the value-added time series classification and clustering procedure can be turned into a software product to classify and re-classify sales data for companies to order better and deliver. Business insights on clusters and predictions can be incorporated as well. Especially for companies with limited labor resources to make predictions on plenty of products, this approach can still generate a good prognosis for all the series. Likewise, the reconciliation with neural network (R-NN) can be used to forecast many time series simultaneously consistently. Then a coherent hierarchy of forecasts can assist subsequent supply chain decisions such as production and deployment.--Author's abstract