Statistical Analysis/Prediction

Situation and Challenge

It is not unusual for one leading player in the Swedish grocery industry to administer tens of millions of time series representing the company’s range of items, distributed over hundreds of stores and some ten warehouses. Over time, the player has acquired expertise in product sourcing, logistics and handling products in a central organization. The central organization supplies stores, which are generally privately owned, with goods based on an inventory requirement as specified by the store owner.

Appropriate stockholding is essential in terms of ensuring profitability, especially in the highly competitive environment of the grocery industry.

From a long-term perspective, stores and warehouses with inaccurate stockholding will cost the player money as a result of lost sales (low forecast) or tied up capital (high forecast).

Required skills – problems with implementation

The customer had previously shown interest in having a central application, with responsibility and functionality, to provide the entire company with forecasting for stockholding. The aim and objective of the application is to execute data-driven, objective forecasts with short lead times to reflect stockholding on the basis of actual demand on item level. There is also interest in reducing the burden on the individual store owner or central stock planner. The objective is to streamline and optimize the number of items that reach warehouses and ultimately stores.

What is a structurally correct approach?

To achieve an automated and data-driven forecasting application that supplies reasonable forecasts for tens of millions of unique time series, good interaction and infrastructure are required between support systems such as data warehouses, forecasting engines, purchasing systems and central warehouse systems. Well-coordinated interaction between these systems is fundamental for success, together with clear and well-supported requirements formulated by knowledgeable employees in consultation with statisticians, developers and testers.

The task of forecasting tens of millions of unique time series in an automated and data-driven manner is complex by nature. The complexity in our assignment can be attributed to the diversity of the time series and items managed by the application. Developing a perfect application from the start can seem daunting. We believe that an application that is robustly developed with sensible architecture can be developed over time based on need.[C.B1]

We have done this before

Our substantial expertise in software, primarily within analytical and technical spheres, allowed us to help the customer by using SAS and several support modules that were adapted for large-scale applications with data on terabyte level.

The analytical sphere:

In this area, we shared our knowledge within SAS/ETS, SAS/STAT and SAS/HPF, together with corresponding theoretical knowledge, which resulted in successful and recurrent releases where the forecasting quality of the application is improved.

The analytical sphere concerns time series trends, time series levels, seasonal time series, time series with a very short history, time series with a sales break, low volume time series which are subsequently combined with special occasion forecasts, historical campaigns and historical empty shelves.

We provided basic concept analyses where ideas and hypotheses were implemented and tested on full scale in development environments to demonstrate the beneficial effects before a final solution was put into production.

The technical sphere:

In this area, we shared our knowledge within SAS/ACCESS for Oracle, SAS/CONNECT and SPDS to manage data supply, multi-threaded environments and quick data management locally in SAS.

This project allowed us to contribute greatly to a small data group’s success with an application that now supplies all of the chain’s stores with item forecasts and purchasing support for product planners who work centrally in the organization.