Time Series Models to Forecast Mail Volume

• Activity Number: 256 - Contributed Poster Presentations: Section on Statistical Learning and Data Science
• Type: Contributed
• Date/Time: Monday, July 29, 2019 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Xuemei Pan* and Mary Pritts
• Companies: and IBM
• Keywords: Time Series Models ; Forecasting; seasonality ; Exponential Smoothing; linear regression; ARIMA

Abstract

Forecasting has become a hot topic in many areas recently. In this application, various time series approaches were developed by the Advanced Analytics team to forecast US Postal Service mail volume across different mail products. Our objective was to provide an accurate and reliable estimate of future mail volumes using historical seasonality and linear trends. The actual mail volume is available at the end of each period; hence we have the correct mail sizes to measure the accuracy of the models. Models including simple linear regression, Autoregressive Integrated Moving Average (ARIMA), Unobserved Components (UCM), and Exponential Smoothing – Winter’s Additive were developed and tested using historical mail volume data. Moving-average and autoregressive parameters were populated; residual correlation, white noise and other model fit metrics were checked against different models. The best model has been selected and applied in real practice.