______________________________________________________PROCEEDING OF THE 25TH CONFERENCE OF FRUCT ASSOCIATION

Data-Driven Approach for Dynamic Pricing

for Decision Making Systems in Marketing and
Finance
Petr Gladilin, Irek Saitov
ITMO University
Saint-Petersburg, Russia
[email protected], [email protected]

Abstract—Recently, the use of accumulated data in the trade profitable market segments. The era of big data challenges new
and financial sector has become increasingly important to tasks associated with efficient aggregation, processing and
improve the quality of services provided. For successful business, interpretation of large amounts of weakly formalized data. The
decision-making systems in retail should combine various main tasks closely related to the analysis of data in the socio-
marketing strategies with analysis of available data. For suppliers
of goods and services and banks acquirers the forecast of the
economic sphere are associated with predictive analytics and
success of outlets and customer's churn prediction are crucial, forecasting in retail, performing optimal marketing strategies,
especially when price policy changes. We propose data-driven customer’s churn prediction models etc. [2], [6].
approach that evaluates clusters of the stores based on the success One of the tasks of price discrimination is to set prices that
of each segment of suppliers and develop the model for assigning
the optimal bank’s acquiring rate within the shopping center
will allow you to sell the product to those categories of
taking into account different customer's loyalty rate accordingly customers who would not buy it at higher prices. And, at the
to their success. In this study we analyze the data on consumer's same time, increase them for segments with high purchasing
activity inside large shopping centers on the basis of the power. Using data-driven approach it is possible to analyze the
transactional dataset provided by our partner bank. We highlight dynamics of sales, comparing it with statistics accumulated
the segments of outlets based on their location, category of goods over the previous time, and to take into account many factors
and services, average monthly turnover and the number of in pricing. But the idea remains the same - to maximize
transactions and apply the constructed model to recommend the company profits.
optimal rate. The proposed approach can be applied in the retail
sector by large suppliers of goods and services and banks. Most dynamic pricing models involve complex
mathematical calculations and computer programming [5].
I. INTRODUCTION The basis of dynamic pricing is the use of information about
As it states from World Payments Report 2018 in the last various characteristics of customers (outlets) and related
year, the volume of non-cash transactions almost doubled the market conditions. Based on this information, the optimal
number of cash withdrawal transactions from bank cards [1]. interest rate for each client group selected using cluster
This indirectly confirms the statement, which analysts have analysis is established, which provides the maximum income
been talking about for a long time that money becomes taking into account the probabilities of an outflow of
“electronic”. customers from target groups. If the margin is too high, they
would go to the competitor, and in case of too low margin,
However, given the ever-increasing importance of they could pay more for the same offer.
electronic data for business in e-commerce, the significance of
using in-depth knowledge of market features and the The main objective of this study is to develop a data-based
characteristics of its main players - customers, suppliers and approach to optimize marketing strategies and bank acquiring
competitors - also in-creases. In order to sale best business systems by the example of shopping centers. The rest of the
marketers should analyze what market segment has the highest paper is organized as follows. Section II presents related work
willingness to buy and to pay and what specific price level will on methods which were applied to solve similar problems
generate the highest market share, revenue and profits [2]. earlier. Section III gives formal description of the available
Detailed segmentation based on socioeconomics, value drivers, data on which we made extrapolation of the data and further
decision drivers and loyalty to pay more generates the most modeling, Section IV presents the mathematical formalization
important data about the market [2], [4]. of the proposed dynamic loyalty model. Section V is for the
methodic of experimental study and experimental results. And
The modern approach to the differentiated pricing strategy finally, Section VI presents conclusion and discussion.
involves wider use of the available information that could be
collected from the retailers for the further assigning of different II. RELATED WORK
prices for the same goods and services in relation to different
categories of customers or at different times [2], [3]. Such The dependence of consumer sensitivity to price changes
strategies allow suppliers to regulate demand throughout the was discussed in [4], [6], [7] on the basis of historicaldataon
year (weeks, days) and enable them to work with the most relations between suppliers DQG FRQVXPHUV

ISSN 2305-7254

Authorized licensed use limited to: KIIT University. Downloaded on January 14,2025 at 16:17:50 UTC from IEEE Xplore. Restrictions apply.
______________________________________________________PROCEEDING OF THE 25TH CONFERENCE OF FRUCT ASSOCIATION

Fig. 1. Distribution of the number of the stores in different categories: left: in total for Northwest region of Russia, right: for 10 largest shopping centers in St.
Petersburg

Main ideas of the customer’s satisfaction analytics can be proposed model is based on aggregated transactions dataset
summarized as follows: (i) the longer the relationship, the provided by our bank-partner, supplemented by data from the
higher the propensity to use a certain product of the company; web-sites of banks-acquirers. Such data in most cases is the
(ii) the longer the relationship, the less sensitivity to price only thing analysts can work with, therefore the development
changes for a certain product of the company; (iii) the greater of such a decision-making model is a popular business task.
the Relationship Breadth (RB, the amount of goods delivered
to the customer), the higher the propensity to use a particular III. DATA AND MODELING
product of the company, but depending on the duration of the
relationship; (iv) the greater the RB, the greater the sensitivity A. Data
to price changes for a specific product. Nevertheless, 1) Shopping centers characteristics
historical data on customer interactions are often either
We use the dataset provided by our bank-partner (one of
inaccessible or insufficient to provide a comprehensive
the largest regional banks in Russia) to extract the information
analysis and identify the degree of customer satisfaction.
about electronic purchases in 10 largest shopping malls T(i),
Providing special marketing research or special surveys is also
i={1,…10} in Saint-Petersburg during the last 5 years. The
often not applicable, therefore, it is necessary to use an
binding of transactions to a specific shopping center was made
available data to form the optimal marketing strategies.
on the basis of the addresses of the shopping center and the
The degree of customer satisfaction is directly related to the geo-location of the successful transaction. The data was
price elasticity of this customer segment [5]. Many studies extracted from the following 10 shopping centers: “Galeria”,
investigating customer’s churn and the dependence of “Europolis”, “Mercury”, “MEGA Dybenko”, “MEGA
customer’s satisfaction on the existing pricing policy Parnas”, “Piterland”, “LETO”, “City Mall”, “Continent
(including discussed above) used data accumulated over a Zvezdnaya” and “Nevskiy Center”. Given information about
large historical period [4], [7], [8]. This made it possible to the purchases in different stores O(ij), i={1,…10}, j={1,….N(i)}
investigate the response of various customers to the changes in from the specific shopping center T(i) all the transactions were
pricing strategy or targeted offers, solving the classification or merged into 18 groups accordingly to the Merchant Category
regression problem [6]. Codes (MCC), which were assigned to the category of each
transaction.
In this study we propose an approach dealing with customer
churn and dynamic pricing strategy in the case where complete Fig. 1 represents the distributions of the number of the
historical data on the interaction of retail outlets with banks (or stores in specific category summarizing the data: (i) in total for
any suppliers of goods and services) is not available. The every transaction in the dataset (to the left) and (ii) only for the

---------------------------------------------------------------------------- 103 ----------------------------------------------------------------------------

Authorized licensed use limited to: KIIT University. Downloaded on January 14,2025 at 16:17:50 UTC from IEEE Xplore. Restrictions apply.
______________________________________________________PROCEEDING OF THE 25TH CONFERENCE OF FRUCT ASSOCIATION

extracted transaction in the largest shopping centers of Saint- To evaluate monthly turnover for the specific store O(j) in
Petersburg (to the right). One can see that the most popular (in the shopping center T(i) we have taken into account
terms of the number of stores) categories significantly differs information about the market share for our bank partner, and
in these two cases: in shopping centers, the most popular have multiplied the cumulative turnover value for the store in
categories are “Clothing, shoes & accessories” stores, the specific category in our data by this coefficient. The result
“Restaurants”, and “Children's wear”, while in the whole of evaluation shows that 96% of the outlets in the shopping
dataset markets mostly correspond to the “Food” category. centers have relatively large turnover (with > 300 000 rubles
per month).
We added new values to the existing characteristics of
transactions in shopping centers: the distance from the nearest
metro station, reflecting the geo-availability of the outlet, the
average number of transactions in a particular category in the
store, and the average monthly transaction value in a given
category.
For clustering and highlighting segments of retail
categories in shopping centers, we have used data with
extracted set of 62 “category in the mall” groups of retail
stores in 10 shopping centers. Next, using implementation of
the DBSCAN algorithm by scikit-learn Python library, 5
groups of stores were identified. 2D representation of these
clusters made by the tSNE algorithm is shown in Fig. 3.
Taking into account marketing research described above
(see discussions in [3], [4], [6], [7]), we have marked the
segment with a high average monthly turnover and top
categories (big black markers on the Fig. 3) as a segment with Fig. 2. The distribution of acquiring rates of 20 largest Russian banks for retail
greater loyalty to increasing the rate of acquiring, and the stores in dependence on its monthly turnover. Data was obtained from public
sources and official websites of banks
remaining groups as a segment with reduced loyalty
(“Others”).
The study [8] proposes an approach that relates tolerance
B. MODELING to price change with an indicator of customer satisfaction. And
since customer satisfaction in the case of store-bank
1) Calculation of acquiring rates interaction directly depends on sales, it is reasonable to
The rate of payment (authorization) consists of: associate the value of the acquiring rate with the turnover of
interchange fee (commission re-turned to the bank that issued the point of sale, which we observe in bank’s practice and in
the card), commissions charged by International Payment the experimental results of clustering in Section III.
Systems and the mark-up of the acquiring bank and. Thus the 2) The model of dynamic loyalty
bank’s revenue is determined only by the last term. As the rate
of payment and interchange fee can vary a lot depending on We develop the mathematical model to evaluate the optimal
the bank’s marketing strategy and existing agreements rate for the specific segment of retailers in the shopping center.
between banks in this study we will evaluate full rate of This model must meet the following conditions: (i) acquiring
payment which should pay the store to the bank-acquirer for rate should be different for retail outlets from different network
each transaction. retail segments; (ii) the probability of leaving for the customer
(point of sale) increases with an increase of the acquiring rate;
Russian banks have different strategies for assigning the (iii) the loyalty of retailer’s outlets to the increase of the
acquiring rate and very few of them take into account the acquiring rate differs for different segments.
categories of goods and services which are of the point of In order to distinguish between the segment of loyalty stores
interest for the specific store. The most popular characteristic (customers of the bank acquiring system) and the segment of
which affects the acquiring rate as the bank offer for retailers customers who violently reacting to any changes in pricing
is the monthly turnover of the specific store. In the Fig. 2 we policy we introduce the loyalty rate ݇ௌ . This rate should be
collect the open sources data from the Top-20 largest Russian greater than 0, and the case when ݇ௌ is close to 0 corresponds
bank’s web-sites about the dependence of the acquiring rates the situation when the segment is intolerant for the price price
on the turnover of specific store. It can be seen that there is a changes, and will almost certainly go away when it rises. Bank’s
clear tendency to a decrease in the rate with an increase in proceeds from the store O(j) from the shopping center T(i) will
thus depend on the loyalty rate ݇ௌ , which corresponds to this
ͳ ሺ௜ሻ (1)
ܲሺ௜ሻ ሺ‫ݎ‬ǡ ݇ௌ ǡ ‫݌‬ௌ ሻ ൌ ‫݌‬ௌ ൅ ሺ‫ ݎ‬െ  ‫ݎ‬଴ ሻ store. The probability of the store’s churn P(i) for the specific
݇ௌ loyalty segment S is:
monthly turnover. where r0 – current acquiring rate for the store O(i), r – chosen
2) Stores segmentation acquiring rate, ‫݌‬ௌ – is the initial probability of the outlet’s churn.

---------------------------------------------------------------------------- 104 ----------------------------------------------------------------------------

Authorized licensed use limited to: KIIT University. Downloaded on January 14,2025 at 16:17:50 UTC from IEEE Xplore. Restrictions apply.
______________________________________________________PROCEEDING OF THE 25TH CONFERENCE OF FRUCT ASSOCIATION

almost from the categories “Clothes, shoes & accessories”,

“Children’s wear”, “Restaurants” and “Sports”.
Thus, we propose the following algorithm for evaluating the
optimal acquiring rate for the specific segment S of customers
reflecting main ideas of the described model of dynamic
loyalty:

Algorithm 1 Algorithm for calculating the optimal

acquiring rate for the customers segment S
ሺ௜ሻ
Input: ‫ݎ‬଴ ՚ current acquiring rate for the category of stores,
‫݌‬ௌ ՚ initial probability of the outlet’s churn,
݇ௌ ՚ loyalty rate,
R ՚ array of rates,
f ՚ income from one store in this category in the
shopping center,
n ՚ number of stores of this segment
For r in R:
ሺ௜ሻ
‫ܨ‬ௌ ՚ ݊ ൈ ‫ ݎ‬ൈ ݂
ଵ ሺ௜ሻ
ܲሺ௜ሻ ՚ ‫݌‬ௌ ൅ ሺ‫ ݎ‬െ  ‫ݎ‬଴ ሻ
௞ೄ
Fig. 3. Experimental result of DBSCAN clustering algorithm and further tSNE
2D projection of store’s segments in ten shopping centers. Each point on the For j in range(0, n):
graph correspond to the “category in the mall” group and may include p ՚ random(0, 1)
different number of the stores. The graph clearly distinguish the segments of If p < ܲሺ௜ሻ :
the stores according mostly to their category and monthly turnover. On the ሺ௜ሻ ሺ௜ሻ
graph, small markers indicate groups of outlets with a potentially low loyalty ‫ܨ‬ௌ ՚  ‫ܨ‬ௌ െ ‫ ݎ‬ൈf
factor (low turnover and a small number of transactions), large markers ሺ௜ሻ ሺ௜ሻ
‫ݎ‬௢௣௧ ՚ ܽ‫šƒ ݃ݎ‬௥ ሺ‫ܨ‬ௌ ሻ
indicate outlets with a potentially high loyalty factor (high turnover and a
ሺ௜ሻ
large number of transactions) Ouput: ‫ݎ‬௢௣௧
Thus, the optimal rate for a specific segment should
maximize the bank income and can be evaluated as Fig. 4 shows the dynamics of the normalized revenue
follows: depending on the increasing acquiring rate for the three basic
cases: loyal customers, neutral customers and disaffected
ሺ௜ሻ ሺ௜ሻ (2) customers. Loyal customers are positive and loyal to raising
‫ݎ‬௢௣௧ ሺ݇ௌ ǡ ‫݌‬ௌ ሻ ൌ ܽ‫šƒ ݃ݎ‬ሺ‫ܨ‬ௌ ሺ‫ݎ‬ǡ ݇ௌ ǡ ‫݌‬ௌ ሻሻ
௥ prices - this category of customers is most suitable for
where ‫ݎ‬௢௣௧ െ is the optimal acquiring rate for the segment S applying the approach described here. The second category of
neutral customers has practically no effect (up to a point) on
and FS(i)(r,݇ௌ ,‫݌‬ௌ ) – is the sum of bank incomes from the
revenue - for them, revenue growth is exactly offset by an
customers of the segment S.
outflow of customers due to price increases. And finally, the
ሺ௜ሻ third category - disaffected or frustrated customers - for them,
The maximally achievable value of the bank’s revenue ‫ܨ‬௠௔௫
from the shopping center T(i) should be evaluated as any increase in the rate will lead to their refusal of service and
follows: transfer to a competitor.

ሺ௜ሻ ሺ௜ሻ
‫ܨ‬௠௔௫ ൌ  ෍ ‫ܨ‬௦ ሺ‫ݎ‬௢௣௧ ǡ ݇ௌ ǡ ‫݌‬ௌ ሻ (3)
௦

The model formulation (1) reflects the idea that even with no
change in acquiring rate there is still the probability ‫݌‬ௌ of
customer’s churn and this value increases proportionally to the
change of the acquiring rate r and depends on the loyalty rate
݇ௌ . We propose to split all the stores into two segments
accordingly to the loyalty rate which is ruled by success of the
store, as we have done in Section III (A): “Segment 1” consists
of non-loyalty customers (stores with low turnover and low
frequencies of transactions) and “Segment 2” consists of Fig.4. Revenue dynamics due to the acquiring system for three customer
successful, and therefore, potentially loyalty stores, which are segments with different loyalty rate

---------------------------------------------------------------------------- 105 ----------------------------------------------------------------------------

Authorized licensed use limited to: KIIT University. Downloaded on January 14,2025 at 16:17:50 UTC from IEEE Xplore. Restrictions apply.
______________________________________________________PROCEEDING OF THE 25TH CONFERENCE OF FRUCT ASSOCIATION

Fig. 5. Experimental results of the probabilistic model for bank's income (relative values) from the acquiring system implemented in 10 largest shopping centers in
St. Petersburg, depending on the cost of acquiring for two segments of stores-customers with different loyalty rate to price changes. Bold line corresponds to the
loyal segment of the stores and dashed line is for the non-loyal ones

---------------------------------------------------------------------------- 106 ----------------------------------------------------------------------------

Authorized licensed use limited to: KIIT University. Downloaded on January 14,2025 at 16:17:50 UTC from IEEE Xplore. Restrictions apply.
______________________________________________________PROCEEDING OF THE 25TH CONFERENCE OF FRUCT ASSOCIATION

IV. RESULTS AND DISCUSSION V. CONCLUSION

We have simulated the change in the bank's income In this study we propose an approach for stratified
FS(i)(r,݇ௌ ǡ ‫݌‬ௌ ), from the acquiring system in the shopping assignment of acquiring rates for the specific stores in
center, subject to a price increase within the selected range shopping centers. The model of dynamic loyalty takes into ac-
from the two proposed in the Section III loyalty segments of count the data about the distribution of the bank acquiring
the stores. rates for retail outlets with a certain turnover and calculates
Fig. 4 represents experimental results of the modeling set how much the bank can raise the rate, considering the
for 10 shopping centers regarding increase of the acquiring prospective customer loyalty.
rate starting from the average rate r = 1.8% for the stores with We have demonstrated that the weighted average turnover
turnover greater than 300 000 rubles per month (see Fig. 3). is the most important feature of the outlet for being
distinguished from the others. However, average transaction
We have assigned k1 = 0.50 and k2 = 0.38 for the loyal and
numbers and the other characteristics of the store (like the
non-loyal segments respectively. The value k1 = 0.50 is chosen
distance from the underground station, average transaction rate
in such a way that the model reflects the case when the
for the mall, etc.) are not an important factor for the
maximum rate in the market is reached (for the selected range
segmentation. This fact confirms the correctness of the banks
of values, it is rmax = 2.5%), the outflow of customers begin to
approach to assign the rate of acquiring based on the turnover
suppress the increase in revenue from the rate increase. Such a
of the store.
choice of k1 guarantees the presence of a local maximum of
revenue corresponding to the optimal interest rate. The value This approach has natural limitations associated with the
k2 = 0.38 for this data corresponds to the threshold for the need for a sufficient amount of data on outlets for clustering
model, which states that non-loyalty segment almost certainly and intelligent selection of a segment to determine the optimal
leaves the supplier as the price increases. Such a behavior one bidding strategy. In addition, the data approach is applicable
can see on every chart on the Fig. 5 (dashed line). These for sufficiently large network retailers and banks, for which
coefficients are analogous to well-known price elasticity there are a large number of customers, which is necessary
coefficients, taking into account possible different customer from a statistical point of view for the adequate functioning of
behavior [9]. A distinctive feature of our approach is that these the model. The model described in Section III can be
coefficients are estimated not by special marketing research, generalized to the case of a possible price reduction while
but by analyzing the available data. taking into account the likely influx of new customers. This
approach requires additional market research and analysis of
Following the results of the clustering analysis of various
related areas with similar product categories of outlets served
customer segments carried out in study [6] the probabilities of
by competitors to calculate the likelihood of their transition to
the initial outlet’s churn ‫݌‬ௌ were equated to 0.05 for non-loyal service at a better price.
(s = 2) and to 0.04 for the loyal segment (s=1). Ten charts in
the Fig. 5 show experimental results of an average of 500 To increase the accuracy of this approach it seems
epochs of the evaluating bank income by Eq. 1 with random appropriate further to collect additional information about the
choice of the acquiring rate r within the range corresponding to state of the business, customer reviews, financial reporting in
the large turnover group. In every step of the rise of the the public domain, etc. Results discussed in this study may be
acquiring rate r the bank income FS(i)(r,݇ௌ ǡ ‫݌‬ௌ ) tend to increase useful in bank Research & Development Departments in order
in accordance with Eq.1 but at the same time, the probability to improve the services and increase the revenue of the bank as
ܲሺ௜ሻ of customer outflow and loss of profit increases. Thus, well as auditors engaged in business analytics and planning the
depending on the loyalty rate, the dependence of the acquirer's optimal development strategy for the customer companies. But
bank's profit on the rate when it is raised will have a convex it is necessary as well to take into account the price of profits
shape with a maximum corresponding to the optimal profit- increase due to this approach - in terms of data mining cost,
maximizing acquiring rate ropt(i)(݇ௌ ǡ ‫݌‬ௌ )At the same time, with a software developing and so on.
greatly increased or greatly reduced loyalty rate, the graph will The proposed approach is applicable not only for adjusting
degenerate, respectively, into a monotonically increasing or the optimal interest rate for acquiring a bank, but also for
monotonically falling function. network retailers with a large customer base and outlets with
The jitter on some charts is caused by a small number of accumulated data on attendance and sales for each of them.
categories of outlets rep-resented in this shopping center, so The method of clustering and subsequent pricing of the
the likely churn of one of the customer-stores greatly affects product, maximizing profits from this group of customers,
the total income in this case. taking into account the possible outflow, can be effectively
used in marketing analytics and building an optimal business
Experimental results on the Fig. 5 show that for the specific strategy for a network retailer.
set of shopping centers (as “Galeria”, “Leto”, “Europolis”,
“Mega Dybenko”) competent choice of acquiring rate can ACKNOWLEDGMENT
increase the revenue from interaction with the loyal customers
segment by 2-3%, while the rate itself will change by five This work financially supported by Ministry of Education
tenths of a percentage point. By analyzing such charts, bank and Science of the Russian Federation, Agreement
analysts or marketers can evaluate the allowable range of price #14.575.21.0165 (26/09/2017). Unique Identification
changes (interest rates) depending on the goals of the business. RFMEFI57517X0165.

---------------------------------------------------------------------------- 107 ----------------------------------------------------------------------------

Authorized licensed use limited to: KIIT University. Downloaded on January 14,2025 at 16:17:50 UTC from IEEE Xplore. Restrictions apply.
______________________________________________________PROCEEDING OF THE 25TH CONFERENCE OF FRUCT ASSOCIATION

REFERENCES [5] P. Kotler, Marketing Management, sixth edition, New Jersey,

Prentice-Hall Inc., 1988.
[1] Official website of The World Payments Report, Web: [6] A.C. Bahnsen, D. Aouada and B. Ottersen, “A novel cost-sensitive
https://worldpaymentsreport.com/ framework for customer churn predictive modeling”, Decision
[2] D.B. Orekhov, “Economic model of the shopping mall: marketing Analytics, vol.2, June 2015, pp. 1-15.
aspects. Marketing problems, logistics”, Problems of the modern [7] E.W. Anderson, “Customer Satisfaction and Price Tolerance”,
economy, vol.4(52), 2014, pp. 181-184. (in russian) Marketing Letters, vol.7, July 1996, pp. 265-273.
[3] K.O. Bochenina, A.V. Boukhanovsky and J.A. Holyst, “Financial [8] J. Dawels, “The Effect of Service Price Increases on Customer
information technology decision support systems based on big data”, Retention”, Journal of Service Research, vol.11, Feb.2009 pp. 232-
Control Engineering, vol. 3(81), 2019, pp. 20-24. (in russian) 245.
[4] A. Tamaddoni, S. Stakhovych and M. Ewing, “Comparing Churn [9] R.H. Frank, Microeconomics and Behavior, seventh edition, New
Prediction Techniques and Assessing Their Performance: A York, McGraw-Hill, 2008.
Contingent Perspective”, Journal of Service Research, vol.19(2), [10] T. H. A. Bijmolt, H. van Heerde and R. Pieters, “New Empirical
Nov.2016, pp. 123-141. Generalizations on the Determinants of Price Elasticity”, Journal of
Marketing Research, vol.42, Apr.2005, pp. 141-156.

---------------------------------------------------------------------------- 108 ----------------------------------------------------------------------------

Authorized licensed use limited to: KIIT University. Downloaded on January 14,2025 at 16:17:50 UTC from IEEE Xplore. Restrictions apply.