Data Science

Predict and better manage returns

More and more customers are shopping online. More and more items are being returned.


According to the Retourentacho of the University of Bamberg, around 280 million parcels and 487 million articles were returned in Germany in 2018. This corresponds to around one in six parcels delivered and one in eight items ordered. Despite all efforts, returns are simply part and parcel of online retailing. This is particularly true of the fashion industry. The increasing number of items and packages that are returned poses major challenges for online retailers and especially for the returns process. From goods acceptance, inspection and classification according to recyclability, to reprocessing and storage of resalable items in the new goods inventory or disposal of non-resalable products, to customer communication and refund of the purchase price: returns processing is a complex and cost-intensive process. Therefore, it must be constantly improved and optimized. But how? To speed up returns handling, capacity planning based on forecasts of expected returns is needed. The better the forecasts, the more efficiently returns can be processed.

Is your warehouse often overflowing with returns? Do you sometimes have too many and sometimes too few staff to process the number of parcels? With this article we will show you how you can avoid exactly these scenarios by using Machine Learning and the prediction of return quantities.


Online retail is currently jumping from sales high to sales high and further growth is expected in the future. However, this also means an increasing number of returns. Returns present online retailers with major financial, logistical and organizational challenges. The primary goal of online retailers is therefore to avoid returns through preventive measures (preventive returns management). The options here range from classic measures such as detailed product information or the display of genuine customer reviews to more innovative approaches such as online consulting services in the form of virtual dressing tools or the use of machine learning algorithms for personalized product recommendations or the determination of customer-specific return probabilities.

Regardless of the efforts, returns still occur. According to a study by EHI, the average return rate is around 20%. However, there are enormous differences depending on the product category: While less than 10% of food and beverages are returned, for example, the fashion and accessories category is the clear front-runner with almost 40%. Returns are therefore simply an unavoidable part of online retailing.

For online retailers, high return rates mean high costs. The average processing costs of a return are €7.93. The drivers here are transport costs for return shipping and personnel and material costs for the returns process. In addition, there are further return-related costs, e.g., due to a loss in value of the goods or for call center personnel. The costs vary greatly depending on the number of returns, the sector and the item, as shown in the table below.

Average process costs of a return as a function of the number of returns (Source: Asdecker (2017). Statistics Returns Germany - Definition)


Due to the increasing number of items and packages that are returned, returns thus represent an enormous cost factor that needs to be controlled. The aim is to process returns quickly and efficiently in order to, among other things, reduce the costs per return and keep the lead time from receipt of goods to readiness for resale as short as possible. Effective curative returns management thus begins with planning the quantity of returns. This requires forecasting the expected number of returns and packages. With the availability of more and more data and machine learning technologies, there are entirely new opportunities to create the most accurate returns forecasts possible.


Returns forecasting is about predicting the quantity and timing of returns. For the planning of logistics processes and resource scheduling, it is particularly necessary to predict the expected returns at the parcel level. This is because, in order to enable efficient returns handling, personnel is needed. However, the number of employees required per day or per shift depends heavily on the number of parcels to be processed. The returns forecast thus provides answers to questions such as: What volume of returns, and thus parcels, can I expect tomorrow, next week and in the weeks to come? Under what circumstances will an order be returned, what are the most important return drivers? And how many items are in a package on average?

The better the future volume of returns can be predicted, the better staff deployment can be planned and adjusted to fluctuating capacities. This is important for companies to be able to process returns efficiently and quickly. By minimizing lead times, items can be brought back online more quickly and are available for new customers to purchase. Ultimately, not only can sales be increased, but the customer also benefits through a quick refund. Ultimately, this also contributes to increased customer satisfaction and loyalty. On the other hand, forecasting the quantity of returns naturally also enables targeted control of either internal logistics resources or corresponding logistics service providers.

The returns forecast therefore has great potential: it enables you to get a clear overview of the associated costs and, above all, the required capacities for logistics, warehouse and personnel. In this way, the returns forecast not only ensures that your warehouse does not descend into chaos, but also supports your financial planning. But how does it all work? And how does machine learning come into play here and what data does it need?


The goal of returns forecasting is to train machine learning algorithms to recognize patterns based on historical returns that allow conclusions to be drawn about the quantity and timing of future returns.

Machine Learning (ML) is a sub-discipline of Artificial Intelligence (AI) and a generic term for artificially generating knowledge from experience. It involves training ML algorithms to automatically identify patterns and relationships in historical data. These identified patterns can then be applied to new data to make predictions.

So for returns prediction to succeed, a sufficiently large data history of returns must first be available. However, due to the naturally high level of digitization, online retail in particular has far more data that can be used to create a returns forecast that is as precise as possible. This includes the following, among others:

  • Number of orders
  • Information about the product sold (e.g. size, color, manufacturer, price)
  • Information about the customer (e.g. gender, age, place of residence, return behavior)
  • Information about the shopping cart (e.g. number of items in different sizes, order total)

Other data sources that can also be used as precursors for returns include information on marketing campaigns that have already been carried out or are planned (e.g., discount promotions and campaigns) or a change in product range or collection (e.g., winter to summer collection).

In addition to these internal data sources, it can also help to include external data, such as the weather, vacations and public holidays, or upcoming major social events (e.g., World Cup), as indicators in the creation of the returns forecast.

Based on this data, it is then possible to forecast how many returned parcels and articles are to be expected and when. The insights gained into the returns drivers not only help to create precise returns forecasts, but are also very valuable in terms of preventive returns management.

With the help of the returns forecast, online retailers thus have a valuable tool at their disposal to manage the necessary evil of returns quickly and efficiently.


Due to the increasing share of online trade and even though online retailers are currently investing a lot in preventive returns management, it can be assumed that the number of returns will continue to rise. Efficient returns processing is therefore indispensable. On the one hand, in order to be able to offer returned items for sale to customers again as quickly as possible (assuming they can be resold) and, on the other hand, in order to meet customer demands for a quick refund for the return or an exchange of the product. Intelligent forecasting is needed to meet this challenge. This is the only way to make optimal use of warehouse capacities and to better plan order quantities and adjust personnel requirements to a fluctuating volume of returns. A returns forecast is therefore a valuable tool for all retail companies that handle their own returns, as well as for returns service providers or logistics companies.

Would you like to know what a returns forecast looks like for a logistics service provider in the online fashion sector? Then take a look at our joint success story with Mode Logistik GmbH & Co. KG, the logistics operator of Fashion ID, the online store of Peek & Cloppenburg KG Düsseldorf.

Similar posts

Bleiben Sie immer auf dem neusten Stand

In unserem Blog dreht es sich um Themen rund um Data Science und KI.