Features, STARTUPS

Big Data for retail

In early February Savings Bank announced launching a neural network. Artificial intelligence will help experts in appraising real estate sites. Meanwhile, retail sector has already learnt to choose premises for new points of sale. BestPlace service allows forecasting a retail point potential by using machine learning algorithms.

Mistake cost

Alexander Kirianov, BestPlace startup founder, likes giving an example of what opening a wrong pint of sale can cost. For Starbucks, the cost of a failure is $500K, while Sears chain faced $60 mio losses following shutdown of its 150 stores. BestPlace helps avoiding such losses, since it helps forecast consumer demand and therefore the rate of revenue at a point of sale, and to subsequently adjust its operation strategy.

BestPlace offers employing the artificial intelligence technology. At the beginning, a neural network is taught on the basis of a big volume of all sorts of data, both open and provided by a customer. Forecasting process is then launched. The service allows assessing potential consumer traffic at a point of sale, number and incomes of the nearby residents, analyzing competitive environment, estimating pedestrian and transport accessibility, appreciating pedestrian and auto traffic routs.

On the basis of analysis of the open data and the data supplied by a customer, machine learning algorithms allow making accurate revenue forecasts”, Alexander Kirianov notes. “Say, an owner expects a new PoS to generate RUR 2 mio a month. Nevertheless, the actual revenue is never above RUR 1.2 mio, so eventually the point of sale is closed having run into substantial losses. With the help of BestPlace, such an owner could spend the very same money to open a similar PoS but in a different location, and then generate the target or even higher revenue”.

Teaching net

Open data which BestPlace operates includes satellite images, maps data services, social networks, real estate market statistics. Much of the data is verified by the BestPlace team through field research and crowdsourcing platforms.

Still, the main asset of the company is the machine learning algorithm which deals with the customer’s corporate data. The algorithm processes data relating to the already operating PoSs (70%) while blind tests are taking place at the remaining points of sale (30%). The system uses the data to learn by analyzing multiple factors that affect demand, and then it selects the most significant ones (some 40 out of 600, for example) which in the aggregate determine demand for a particular brand, for a particular retail trade or services provision format.

A neural network builds thousands of independent expert critical trees which combine factors in different sequences to discover hidden patterns. Then they are processed using the data of the existing points of sale. Ultimately, a system composed of the best trees is formed. Out of thousands of options, only fifty may be selected. That is like choosing fifty best experts who have thoroughly studied a running business and found some concealed correlations”, Alexander Kirianov explains.

The obtained information is first tested by the service itself, then it is verified by the customer at some chosen samples. If the customer sees that some essential demand factors have been found, a contract is made. The selected metrics allow both forecasting the cost advantages of a new PoS and monitoring the operating points of sale and their financial viability, since there are incessant changes around every location which are to be predictively responded to.

Benefits

Currently the service has clients representing various industries including fuel stations, horeca, pharmacies, groceries (8 segments in total). The service is used by Gemotest laboratories, and a pilot project with X5 Retail Group is also underway. The startup is advancing to the CIS markets and makes plans of establishing itself in foreign countries.

The main parameter of the service is a possibility to drastically reduce the rate of mistakes in forecasting earning capacity of points of sale. According to the statistics, the rate of mistakes may reach 30% while the service reduces it by one third. The service is provided on a subscription basis and is priced within 5% of the produced economic benefits. For a retail grocery store, for example, 1% mistake probability is priced by the startup at RUR 66K ($1.2K) of lost profit a year.

This predictive service of forecasting revenues of points of sale has a good chance to be successful in case customer base is rapidly expanding and analytical potential is timely improved, Anastasia Arkhipova, market research analyst at CBRE commercial real estate services firm, believes. Retail sales have always operated in a highly competitive environment, she notes. Multiple technological trends have been observed in the industry recently, with development of predictive services being one of those.

In case of this service, the key role is played by the Big Data which includes various and at times incommensurable data and factors for prognosticating consumer behavior thus producing a rather accurate demand forecast and improving operational management efficiency”, Anastasia Arkhipova says.

Choosing locations for the new points of sale remains a challenging issue for the market, Andrei Golubkov, spokesman for Azbuka Vkusa supermarkets, agrees. That is especially true in respect of a chain, as is the case with Azbuka Vkusa, which operates in multiple formats and specific characteristics are valuable for each point of sale.

The idea appears to be interesting and relevant, given the AI broadening use and retail chains expansion, as well as the fact that in a metropolis, performance of points of sale may be affected by most diverse factors one may often be totally unaware of. Predictive analytics on the functioning PoSs is also important, Andrei Golubkov notes, mentioning charges imposed at some point at all parking areas in downtown Moscow which had an immediate effect on retailers’ performance and required a new format to be introduced at many points of sale.

According to Ivan Fedyakov, CEO at INFOLine information agency, such a technology is the next level in the progress of geoinformation services Walmart designed 40 years ago to build up its chain. If a universal algorithm is developed which can be applied to any market player, that would be a good option, he believes. Especially as today even most technologically advanced retail chains quite rarely have appropriate solutions to that challenge.

To open a point of sale today, ten candidate locations have to be reviewed at best. As Ivan Fedyakov notes, last year Х5 Retail Group opened 3,000 new PoSs which therefore required at least 30,000 local researches.

By Olga Blinova

Previous ArticleNext Article