Forecasting demand accurately is key to running an efficient supply chain, keeping inventories stocked to the ideal levels and fulfilling customer orders. However, deciding how much stock to hold is a challenge for hire businesses. There is a fine balance between holding enough stock to fulfill hires and not holding too much stock so that overall utilization is too low to achieve the return on investment. In addition, the number of returned assets also needs to be considered in the forecast to determine how much stock to hold to satisfy demand.
Kaylea Haynes shares a case study on forecasting the demand for thousands of assets across multiple locations. The first approach to the problem used common forecasting techniques such as ARIMA and exponential smoothing, along with the Croston forecasting method for intermittent demand, to forecast the number of new hires and the number of returns in a week. Kaylea and her team split the data into training and evaluation sets and trained all of the models on the training data set. They then used various evaluation measures, based on the point forecast of the mean and the prediction intervals, to choose the method that forecast the best for the particular asset.
Given the scale of the problem and the forecast time horizon required for the case study, the team chose a simpler model. Assuming normality of net demand (returns – hires), they used various percentiles as their forecast. This method required a few data cleaning steps to help get the best forecasts. Initially, they tested for monthly seasonality in the data using a linear regression model with dummy variables for the months. The results of the seasonality then determined which data was used to calculate the forecasts for the next quarter. Next, the team used anomaly detection to remove outliers in the data and changepoint detection to test for any abrupt changes in the distribution, in which they ignored prior data if there had been a change. Using this manipulated data, the team calculated 85%, 90%, and 95% percentiles as their quarterly forecast for the number of assets to have available, on average. The forecast method is a semiautomated process in which the percentile level was chosen manually. To try and get a more automated approach, the team then looked at integer programming methods to optimize the number of assets across the whole fleet.
Kaylea Haynes is a data scientist at Manchester-based data analytics service Peak, which helps companies grow revenue and profits using data and machine learning. Kaylea focuses on developing techniques for demand forecasting. She is a member of the Royal Statistical Society and co-organizes R Ladies Manchester. Kaylea holds a PhD in statistics and operational research from Lancaster University. Her thesis was titled “Detecting Abrupt Changes in Big Data.”
For exhibition and sponsorship opportunities, email strataconf@oreilly.com
For information on trade opportunities with O'Reilly conferences, email partners@oreilly.com
View a complete list of Strata Data Conference contacts
©2018, O’Reilly UK Ltd • (800) 889-8969 or (707) 827-7019 • Monday-Friday 7:30am-5pm PT • All trademarks and registered trademarks appearing on oreilly.com are the property of their respective owners. • confreg@oreilly.com