Within the framework of a proof-of-concept (PoC), the feasibility of predicting return quantities in parcel logistics is investigated. Different approaches, based on both time series analysis and machine learning, are analyzed and benchmarked using appropriate error measures. The goal is to predict the quantity of shipments expected at the shipper's location designated for returns processing in the next three days. PTA advises the customer and identifies appropriate procedures to forecast return quantities.
The selection of the processes is based on the state-of-the-art. In cooperation with the customer, historical data is collected and the benchmark is defined. For this purpose, the time periods for training and test data as well as suitable error measures are agreed upon. Using Jupyter notebooks, the data is first analyzed in terms of a classic time series analysis. Concepts such as stationarity, trend and seasonality of the time series are examined in order to optimally parameterize different forecasting methods such as (S)ARIMA, (T)BATs, Exponential Smoothing, Holt-Winters, Random Forest and XGBoost. The development and evaluation of the prediction models is done using well-known Python libraries such as fbprophet, matplotlib, pandas, scikit-learn, scipy, statsmodels and tbats.
By integrating a forecasting service into the business customer portal, the customer wants to provide its customers (shippers) with reliable information on the quantities of returns to be expected at the respective returns processing locations in the next one to three days. This and the provision of other services is intended to increase customer satisfaction and customer loyalty with the aim of further expanding the returns business.