Time Series Data Mining Through Automatic Forecasting and Decomposition

• Activity Number: 637
• Type: Contributed
• Date/Time: Thursday, August 5, 2010 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Shu-Ngai Yeung*+ and Tom Siu-Tong Au and Guang Qin Ma and William Pepe
• Companies: AT&T Labs and AT&T Labs and AT&T Labs and AT&T Labs
• Keywords: Time Series Data Mining ; Automatic Forecasting ; Effective Business Day ; Seasonal Decomposition ; ARIMA Models ; Exponential Smoothing

Abstract

Due to the advance in IT, industries such as telecom, finance and retails, now are capable of capturing detail information on the usage of service and products. Augmented by lower level data, monthly business performance time series data can be disaggregated temporally (e.g. daily) and vertically (e.g. by products) to uncover business insight and improve forecast accuracy. This paper proposes a time series data mining framework through automatic forecasting for decomposed time series in the enlarging database. We develop an automatic procedure to capture the time varying structure of trend, seasonal, calendar and special-day effects for each time series. Forecasting is done automatically by ensemble exponential smoothing and ARIMA models for each time series component. Simulation results show the improvement of forecasting accuracy of the proposed method over standard methods.