How machine learning predicts demand and sales in retail: Vladimir Kuchkanov, Data Scientist from Competera How machine learning predicts demand and sales in retail: Vladimir Kuchkanov, Data Scientist from Competera

Machine learning in retail is an algorithm able to analyze and predict the demand and sales based on historical network data (the so-called precedents). Machine learning in retail either automates or optimizes any processes related to retail chain operations. Vladimir Kuchkanov from Competera told AI Conference Kyiv about methods of application, solved objectives of large business, and preparations for optimization.

Vladimir Kuchkanov is a Data Scientist at Competera, a company that helps to optimize price formation and increase profit of large retail firms using machine learning. Earlier, he focused on analytics at such FMCG companies as Mars and Philip Morris and was a part of the European committee on price formation at Mars, Inc. Besides, Mr. Kuchkanov gives a Brand Management course at Kyiv Academy of Media Arts.

Given below are the most interesting key points from the expert’s presentation.

Application of artificial intelligence in retail

Kuchkanov divides ML goals into two classes.

The first class of goals is process optimization. AI can work with supply chains: it recommends when and where to order goods, as well as how many of them are required, in order to provide a seamless delivery. AI also allows to optimize expenditures for advertising and marketing in general. Vladimir separates optimization price formation tasks. ML helps to find the best possible price levels for almost all products and shops.

The second class of goals is forecasting. Machine learning allows to predict sales or demand for certain goods. More broadly, the algorithm operation refers to business planning: for example, ML predicts the efficiency of establishing new shops.

ML can give recommendations by processing data on each retail customer (regardless of full names) individually. Therefore, clients will obtain customized special offers, which are much more efficient.

Vladimir believes that it will allow to cooperate with customers via virtual assistants and chatbots in the future.

“Automated commerce is a brand new industry. It allows to foresee the demand for a certain product, as well as to conduct the order, charge money, and automatically deliver goods. But in fact, this technology is not quite futuristic. The research has shown that 80% of the Chinese consumers don’t mind if AI would check the contents of their fridge, promptly make additional orders of products that almost run out, and bring them,” Vladimir Kuchkanov says.

Retail objectives that can be solved by neural network

According to Vladimir, a problem faced by any manager determining price changes is the so-called opacity of the future. In this situation, manager’s solutions have a poor impact on sales. There are four reasons for this issue:

  1. the response of demand for price changes is non-linear: there are certain price caps followed by either rise or drop in sales;
  2. the response of demand is multifaceted: customers respond not only to the product price, but also to the relative price attractiveness within a brand, category of goods, or in comparison with competitors;
  3. the response of demand is asynchronous: the response to the price appears within an average cycle of product purchase. If people buy goods once per month, they will respond to new prices during the month;
  4. the response of demand is lost among other factors: seasonal prevalence, availability of advertising, etc.

Taking into account the above-mentioned aspects, it is quite difficult to determine how the price affects the demand. Neural networks can be useful in this case. In a perfect world, they should perform five types of goals:

  • learn and become smarter day by day;
  • predict and foretell;
  • integrate with retailer’s software; otherwise, machine learning will work poorly, and forecasts will be delayed;
  • solve a maximization problem: what prices should be set for each product so that their total price will be as high as possible;
  • randomly recommend a huge amount of possible prices for an unlimited number of goods.

ML benefits based on Foxtrot case study

The expected outcome of neural network operation is the automatic price control. As an example, Vladimir has presented the result of Competera’s cooperation with a Ukrainian retailer Foxtrot.

To start the test, Foxtrot needed several months to prepare databases. Afterwards, Competera, together with Foxtrot, chose 10 shops where a neural network controlled the prices. Within four weeks, the algorithm was maximizing proceeds without reducing the product marginality.

The neural network considered price and non-price factors: shelf prices, prices in competitors’ stores, information about discounts, seasonal fluctuations, stock balance, and employee engagement, which plays a critical part in price formation.

Results of neural network operations

Then, Competera compared the outcome with 10 shops where Foxtrot used a conventional method in price setting. The company saw the following:

  • revenues in shops applying AI increased by 16% in a month, while revenues in the control group grew by 2.4%;
  • sales in the group of neural network shops increased by 14%, while in the control group by 8%.

Consequently,  the neural network defeated retail experts in terms of key business factors.

“By taking 100% of marginality prior to the test, the control group dropped it by 50% at the beginning of work. It was obvious: the company carried out a promotion. Neural networks did not require such a step, thus, the marginality dropped only by 1.5% during that period. In general, Foxtrot increased prices by 2.8% in a month, while the market stood still,” Vladimir sums up.

What challenges were faced?

Primarily, technical barriers caused challenges. “The company had to spend several months to arrange their data. To make the machine learning algorithm function, they should be in a prefect order,” Vladimir stresses.

Another significant factor is the human fear of new technologies. “Some people said, “I do understand everything, but I’m scared” and delayed the test for a week. Somebody was afraid of pressing the button just because the boss was on vacation. Others stated, “I’ve been working in the retail sector for 20 years, and what about you?””, Vladimir explains.

However, the system was integrated successfully. Currently, Competera is scaling up the neural network to 20 major categories of household appliances at Foxtrot.

What is required to start working with neural network:

  1. clear structured data at least for three years;
  2. readiness of senior executives and team for changes in order not to slow down the process;
  3. your own development department consisting of 6–10 developers or partnership with a technological company.
Related news