How to Predict Demand with the Help of Machine Learning

2K

AI and ML are mentioned with ever-increasing frequency as efficient retail tools. However, implementing and using them properly is a specific task that requires clearly set goals and a professional approach. In this article we’ll talk about implementing ML technology in eCommerce for demand forecasting, what it requires, and how soon can one await the results.

How ML works

The core difference between ML and conventional analysis is that, when tackling a task, a model with ML masters new techniques by processing new information. The model doesn’t follow a strict programmed algorithm. Conventional approaches to analytics are increasingly replaced by ML due to their greater flexibility.

For the model to operate efficiently with an ML algorithm, it is enough to have some historical data.

The advantage of ML algorithms in predicting demand

ML algorithms help to build connections between various data and a multitude of features. Thanks to cloud computing advancement, such solutions become available not only for large corporations but also for small-to-medium businesses. Retailers that have already implemented forecasting with the means of ML methods name the following advantages of ML algorithms:

1. Increased accuracy of a forecast. The forecast accuracy in a product category is 95%, and the average improvement in forecast quality amounts to 15-20 points compared to conventional algorithms.
2. Automated forecast of the volume of goods for sales promotions. ML algorithms give the opportunity to forget about adjusting the volume of goods manually when planning sales promotions.
3. Adaptivity of algorithms. In a situation when demand is unstable due to the coronavirus pandemic, it only takes a week for ML algorithms to adapt to changes in consumption.

Improved accuracy of an automated order system

Low accuracy of an automated order system leads to the surplus of warehouse stock and reduction in liquidity or, on the contrary, to shortages and empty shelves. As a result, customer loyalty declines.

An automated order solution is based on a planned order formula:

Planned order = Sales forecast + Reserve stock – Opening balances

Opening balances are a known quantity, reserve stock is, as a rule, calculated based on the size of a shelf – thus, to increase the automated order accuracy, it is necessary to work with sales or demand forecasting.

Implementation of an automated order system can be divided into two stages:

1. Sales forecast generation with the help of ML algorithms.
2. Automated order system creation (postprocessing according to the above formula).

If a retailer already has an automated order system, then demand forecasting by using ML methods can be integrated into the current system. As a rule, one or two shops and a few categories are chosen for a pilot project.

What do planned processes involve?

• a vision of a future solution is worked out;
• there is an understanding of the basic functionality and improvements;
• a roadmap for project implementation is created;
• the composition of the team is determined: the number of specialists, roles, and dates of involvement;
• the team is provided with tasks for the upcoming sprints, and subsequent ones are gradually being worked out.

Required data

The set of data required for building a demand forecasting model depends on the structure and set of analytical features that a company supports. However, there are minimum requirements. To build an ML model, you need historical data that has been stored for at least two years. Such a requirement is related to the dependence of demand on seasons, hence it is crucial to have data on two reoccurring cycles.

To implement a demand forecasting model, you need:

• Catalogs of goods,
• Catalogs of shops,
• Data on receipts,
• Sales promotion history,
• History of returns, markdowns, write-offs,
• Data on balances

Speaking from the experience of demand forecasting system development, the biggest challenge is to obtain information on sales during sales promotions and markdowns. The storage of such information is often fragmented, and sometimes data on sales promotions is simply not available so sales history during promotions needs to be recreated manually. Sales promotions are a crucial element for building an ML model, that’s why it is recommended to collect and store information about them.

The following data format on sales promotions can be in place:

• Shop ID,
• Item ID,
• Sales promotion starting date,
• Sales promotion end date,
• Sales promotion type (if any),
• Price benchmark,
• Discount rate or promotion price

The model development and testing will take about 1.5-2 months. After the model is built, tested, and the result is evaluated, the decision on its scaling is made.

How an automated order system works

The predictive model is not a turnkey solution by itself but the indispensable core of the system. To introduce an automated order system into a business process, regular data uploading should be established. For this purpose, a service receiving data and transferring it between a retailer’s software, the model, and the cloud data warehouse is built.

If a retailer already has an automated order system, then the forecasting system will transfer data on planned sales to integrated software. If an automated order system needs to be built from scratch, then, according to technical specifications, the information can be submitted to the procurement department in various forms. Advanced cloud technologies and services allow for the implementation of different scenarios – web interfaces with order tables and charts, emailing of order forms, notifications in messenger apps.

Summary

The success of the ML demand forecasting project depends on data quality and proper organization of its collection and storage. On average, after the implementation of ML-powered solutions, the number of markdowns and unscheduled sale events decrease by 20%, while sales during sales promotions increase by 15%. The overall volume of sales grows by 5-10%.

In addition, the implementation of intelligent systems makes it possible to shift routine processes from people to Artificial Intelligence and optimize the staff.

The implementation of ML and AI automates the process and frees up people from the routine processes. Andersen – a software development company – is deeply involved in the automation and the implementation of new technologies into day-to-day processes. They not only develop software for retail companies (which we’ve been successfully doing for over ten years now) but also offer their customers consultations on automation and digitalization.