As an epigraph aka disclaimer, I would like to say that we are planning a series of materials on ML in the service of a modern retailer. We plan to tell "from the stove" to small details (including bolts and screws) about how machine learning saves our business from routine and low margins. We hope that the topic will be of interest to the Habr audience and will not cause an acute allergic reaction among readers. If you have any personal experience on the stated topic, do not hesitate to share it in the comments.
80% of companies are implementing machine learning technologies - this figure was named by one of the leaders of Microsoft Jean-Philippe Courtois at the AI Jorney conference in December, speaking about the impact of the pandemic on the global economy. According to Mr. Courtois, 56% of companies plan to increase their investment in machine learning.
In retail, AI and ML are already used for many operations - from planning supplies and improving marketing efficiency to calculating work schedules for retail employees. M.Video-Eldorado has gone further to offer its customers an assortment, prices and promotions based on advanced analytics. How it works, we will tell in a series of materials about various ML-solutions in the field of commerce.
A few obvious thoughts
To begin with, let's remember what tasks Machine Learning can solve and why it is good for retail. If you sell milk from your cow, you know perfectly well without computers which of your neighbors needs it, how much they will buy and what price they consider attractive, and even here simple accounting skills will not hurt.
It is a different matter - large retail chains with tens or even hundreds of thousands of commodity items and millions of buyers. Which of these will "go to the people" easily and quickly, and which will stand on the shelf for years? What should you still order from suppliers and what should you give up?
What products can you offer holiday discounts, what to offer in contextual advertising? It is impossible to define it “manually” and precisely. And then ML comes to the rescue, machine processing of huge amounts of data, which can be "laid out" by product categories, characteristics, geography of individual points, the speed of sales of goods, etc.
ML algorithms are, of course, not a dogma, but a guide to action for experts taking the final on procurement, pricing and promotions. To put it simply, the "machine" offers a fork based on demand, competitors' prices, and other given parameters.
For example, for one and the same audio system there may be several bids from 2,000 to 4,000 rubles in Moscow and from 1,500 to 3,400 in a regional center with a lower per capita income. If you have a goal to sell faster, you set a lower threshold.
If you want to earn more, do exactly the opposite. And after a certain time, ML tools will tell you whether you did the right thing and whether you need to adjust your actions. Let's say not to chase the maximum profit, but to bet on the rate of turnover of funds.
The virus drives to the digital
Machine Learning capabilities are now used by many, from marketplaces and federal networks to local brands. The migration towards ML became especially noticeable during the pandemic, when business began to massively move "online and digital", which means that much more data appeared for "machine" processing. Retail has become more effective in analyzing the behavior of its customers, their views, searches, participation in promotions, purchases, reactions to communications.
About 74 million users visit the M.Video and Eldorado web sites monthly. Their profile and history of interaction with the company form the basis of predictive models and recommendation services, which the retailer develops, including in the field of commerce.
However, online is no longer the only source of consumer information. M.Video-Eldorado in 2020 completely switched to the OneRetail platform, which, thanks to mobile technologies in retail, allows digitalizing the offline experience of customers and analyzing this data. And this is a huge array - 85% of buyers of equipment in one way or another interact with physical stores.
The seller through the application in his smartphone authorizes the client, gets access to his profile, understands his preferences, sees the purchase history, bonuses and discounts, personal offers. Through this decision, the selection and purchase of goods in the store takes place, which is also added to the analysis and affects future contacts.
Predictive and recommendation systems also analyze sales indicators, level, dynamics of demand, price elasticity, customer engagement and the impact of promotions on sales and business efficiency.
The introduction of data science solutions in commerce will allow M.Video-Eldorado, firstly, to better understand the needs of customers and increase the accuracy of assortment planning, and secondly, to calculate the optimal price based on the desire to make the best offers on the market, while increasing business efficiency.
How to set up an assortment
For example, machine learning helps predict demand not only based on sales already made, but also on customer demand. If you imagine that 12 teapots can be placed on a store shelf, then what models should they be if you have 50 in your assortment?
How to form a stock of a small regional warehouse so that the maximum number of customers receive fast delivery times for their orders? Finally, how do you find the perfect balance between sales growth, market share and business performance?
If earlier these questions were answered expertly by M.Video-Eldorado's commercial managers, now our data-science team is developing ML-based recommendation services to help them.
So, based on user sessions, a decision tree is formed, where all products are grouped based on how often they are viewed together. This allows you to create a balanced assortment and not duplicate products on the shelves that cover one need. Our first story in this series is dedicated to just this CDT.
Determine the right price
M.Video-Eldorado also tests Machine Learning algorithms to create scenarios for automatic calculation of the recommended price and evaluation of the effectiveness of promotions. The goal is to give commercial managers a tool for daily price management based on both internal data (sales volume, margin level, inventory, promotional calendar) and external (market prices, competitor activity, etc.).
The model calculates several scenarios depending on the target indicators and recommends the optimal one. “Now, as part of the pilots, we compare the degree of consumer interest in a particular product in different regions and the level of their sensitivity to price fluctuations,” says Vladimir Litvinyuk, Head of the Competence Center for Applied Data Analysis and Machine Learning, M.Video-Eldorado Group .
It is no secret that when choosing a flagship smartphone or a side-by-side refrigerator, the buyer is looking for where it is more profitable, and when buying a kettle for a summer cottage or replacing a TV in the kitchen, he will prefer a trusted store at home or favorable conditions of the loyalty program.
In addition, the data office team is currently testing algorithms for evaluating the effectiveness of promotional campaigns. We have built a sales forecasting model taking into account a set of stocks and their parameters. Based on this model, various scenarios of sales of a specific product and a category as a whole are simulated for various options for combining promotions, the impact of various sets of promotions on sales growth and the level of cannibalization of promo goods by other goods sold at a regular price is estimated.
In the future, we also expect to learn how to select the optimal price discount and other conditions for the promotion for each specific product, in order to maximize the target turnover or profit from the promotion.
Now the development of Machine Learning in Russia is hampered by two factors: problems with the availability of data for processing and insufficient distribution of ML-models, which was just discussed at the aforementioned AI Jorney conference.
Yes, giants of the level of tech markets, which control up to a quarter of the national market, have a sin to complain about the lack of digitized information. Smaller companies, on the other hand, do not always have a clear idea of what data, in what volume and of what quality are necessary for the successful application of machine learning algorithms in practice.
It is necessary to take into account the descriptions of goods divided into groups and clusters, sales statistics, preferably over a long period, all possible variables: seasonal demand, the schedule of holidays, currency fluctuations, the emergence of new competitors.
In electronics retail, for example, the difficulty with data lies, firstly, in the low frequency of purchases - no one goes for a new TV, refrigerator and headphones twice a week, as for bread and meat. And secondly, the assortment is very diverse and weakly connected, which makes it difficult to find patterns. Nevertheless, digital models must react to the situation in real time: if they did not have time to react, they missed customers.
The question is which platforms to choose for dynamic pricing ... Someone chooses open source, someone proprietary box solutions from vendors, someone cloud ML frameworks ... The number of software tools is measured in dozens, there are options that are free for entrepreneurs. So why do IT conference attendees talk about the lack of proliferation of machine learning models?
The point is again to understand your own needs and assess your capabilities. And here the exchange of experience and best practices is more important than ever. For example, when ML models and dynamic pricing were introduced in the Russian online store BABADU, revenues and profit margins grew by 7% in just a few weeks. The consumer responds to "fair" prices by bringing money to Machine Learning ambassadors.