See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/362549169

Machine Learning for Dynamic Pricing in e-Commerce

Article in Annals of Dunarea de Jos University of Galati Fascicle I Economics and Applied Informatics · December 2021
DOI: 10.35219/eai15840409230

CITATION READS

1 492

1 author:

Maria Cristina Enache

Universitatea Dunarea de Jos Galati
25 PUBLICATIONS 32 CITATIONS

SEE PROFILE

All content following this page was uploaded by Maria Cristina Enache on 17 February 2023.

The user has requested enhancement of the downloaded file.

 Annals of “Dunarea de Jos” University of Galati
Fascicle I. Economics and Applied Informatics
Years XXVII – no3/2021
ISSN-L 1584-0409 ISSN-Online 2344-441X
www.eia.feaa.ugal.ro

DOI https://doi.org/10.35219/eai15840409230

Machine Learning for Dynamic Pricing in e-Commerce

Maria Cristina ENACHE 

A R T I C L E I N F O A B S T R A C T

Article history: Dynamic pricing is a long-term pricing model that can increase the conversion rates of your
Accepted November 2021 e-commerce store. You can use A.I applications to offer different prices for the same
Available online December 2021 product to different customers, depending on unique personal factors. Advanced
JEL Classification applications should take into account many other factors, such as the prices charged by
K22, M21
competitors that buyers have previously sponsored, the current demand for the product,
Keywords: cross-price elasticity, halo ratios, and so on. Some AI-based dynamic pricing models can
dynamic pricing, e-commerce, also implement in-depth learning capabilities to deduce the prices that each customer will
machine learning be willing to pay for a product or service at some point.

© 2021 EAI. All rights reserved.

1. Introduction
In 2021, customizing websites, products, offers, and content recommendation services to help
retailers and other companies customize their proposal to support potential and increase conversions is
becoming mandatory. The opportunities to apply customization are now much wider than the initial retail
recommendations. Personalizing the digital experience is the dynamic delivery of personalized content,
products, or promotional offer recommendations to website visitors or application users based on their
characteristics and intended behavior to support long-term conversion and engagement. Although many of us
will have experienced customization through retail product recommendations, the definition shows the
relevance of personalization for all types of companies. For example, even B2B companies that do not sell
online can use personalization to recommend relevant "best content" offers to support the potential interest
of a business and thus fuel them for a sale.
The limitations of physical interactions around the world have reshaped the lives and habits of
buyers. While the pandemic has affected most industries, e-commerce has thrived. Dynamic pricing helps
retailers refine their pricing strategies to increase profitability and boost customer engagement and loyalty.
For e-merchants, it is vital to keep up with price movements. Overloading your products can lead to
losing customers to your competitors, while overloading can lead to lower revenue. In 2019, online shopping
generated revenue of $ 3.53 trillion for e-merchants. In addition, by 2040, it is expected that approximately
95% of all global purchases will be made online.

Figure 1. Share of e-commerce in total retail sales, UK and USA - 2018-2020 (source [8])
Source OECD


Dunarea de Jos University of Galati, Romania. E-mail address: mpodoleanu@ugal.ro.
 For the United States, data provides estimates for e-commerce as a percent of total retail sales, based
on data from the Monthly Retail Trade Survey and administrative records. Data for the second quarter 2020
are preliminary estimates. For the United Kingdom, data provides Internet sales as a percentage of total retail
sales.

Figure 2. Retail trade in 2019-2020 (source [8])

2. Web customization
When web personalization services first appeared, they were part of expensive analytics solutions
available only to the largest companies, but today there are many services suitable for small businesses.
These are two of the original services that are available today in a new form.
• Adobe Target is one of the most established customization engines evolving from the original Touch
Clarity in 2004 and then integrated into the Omniture analysis suite as Test and Target
• IBM Product Recommendations - The Coremetrics smart offering has been widely used by retailers for
analysis, but since its acquisition by IBM it has been offered as a separate product.
• Google Optimize 360 - This is the latest analytics service offering added to this page's most recent update.
Optimize is available for free as an AB test for all GA users, with some customization features, but the full
version of this customization option is only available to enterprise subscribers today
• Google AI recommendations - Officially launched in January 2020, it is designed to serve retail
recommendations. It's based on Google's experience of applying AI to their search engine.
Recommendations AI allows Google Cloud customers to apply rules to adjust what buyers see and to
diversify the products displayed, filtering by product availability and custom labels. Accepts international
product catalogs in multiple geographic areas and offers recommendations anywhere on a customer's
journey, whether on a homepage, when confirming an order, or in a shopping cart. Once the initial data
import is complete, Recommendations AI users can choose the type of model and the optimization goal.
Model training and adjustment takes between two and five days, says Google, and model
recommendations can be previewed before being offered to customers. Google is introducing a new price
structure with three volume-based price levels for predictions and a separate fee for training and
adjusting the model. All new customers will receive a credit in addition to the free general credit for new
Google Cloud customers, which is usually enough to train a model and test its performance in production
through a two-week A / B test. Google's AI recommendations compete with Amazon Personalize, which
similarly uses machine learning to provide suggestions for websites, SMS, email, and apps. According to
Amazon. Personalize addresses issues such as creating recommendations for new users or products
without historical data through API calls that automate the tasks needed to build, train, adjust, and
implement a recommendation model.

3. Dynamic Pricing using machine learning

Dynamic pricing is the practice of adjusting prices according to market and customer data. E-
commerce business such as eBay and Amazon, leverage different types of dynamic pricing to attract more
customers and increase profitability.

115
 Dynamic pricing is a process of automatically adjusting prices for products or services in real time to
maximize revenue and other indicators of economic performance. To define the optimal price, a dynamic
pricing strategy considers the current state of the market as a basis, including the company's previous price,
changes in competitors' prices, consumer tastes, time frame and other external factors.
Pricing strategies are applied in many business contexts and are used by airlines, railway companies,
concert venues, theaters, car rental companies, accommodation providers and retail companies to respond
promptly to market fluctuations. For example, Amazon monitors and changes the prices of its products every
10 minutes, once big data is updated and processed. Uber also uses a flexible pricing strategy in the event of
high demand caused by bad weather or a particular event pushing prices up.
Any application that has a dynamic pricing component based on machine learning will need a
company's sales and transaction data to calculate demand for each product in your range. This forms the
basis for each AI pricing decision. At least all sales information is required, e.g., what items were sold at what
price.
Every e-commerce business will look for real tangible ways to increase their business success. One
way would be to consider changing the look of the store or hiring a copywriter to write compelling product
descriptions. But a method increasingly used by new e-commerce store owners is the use of artificial
intelligence and dynamic pricing. For some people, the terms "artificial intelligence" and "dynamic price" may
simply be fashionable words without real practical meaning.
You no longer must spend hours looking at competing prices and deposit prices to determine the
best price you should use. Now you can use a combination of machine learning, AI, and algorithms to predict
prices and take care of the most important thing: your customers. This article helps you understand what the
two ideas mean and gives you a framework for starting to implement them in an e-business.
The benefits of using dynamic pricing are increased profits and revenues, as e-commerce business
owners can remain competitive and price their products in line with sales volumes, competition, and market
trends.
Today, in any e-commerce business, many merchants use flexible pricing to stay competitive. Here
are some key benefits of dynamic e-commerce pricing:
• Competitive advantage Automatic monitoring of competitors' prices allows you to quickly adapt to the
dynamic environment and gain market leadership.
• Profits increase After the implementation of dynamic prices; Best Buy recorded a 25% increase in sales.
By analyzing the market, you can adjust the price of a product to generate more revenue. If the demand
for a product is low, you can increase it by lowering the price, and if it is a peak season for a product, you
can increase the price without changing the sales volume.
• Get market information. Ongoing market monitoring allows you to keep abreast of prevailing market
trends and gain information about consumer behavior, which can lead to better decision making.
In e-commerce, retailers generally look at their own sales data as well as external factors to
constantly change their prices when the market fluctuates. But more than that, it is not only an approach that
focuses on price changes, but also on the analysis of competition. The analysis of the competitors starts by
defining them first. Without a doubt, the most popular place to promote an e-commerce business is Google.
Not only is it just a means of promotion, but potential customers are starting to buy there for the first time.
So, it would be a wise start for competition analysis and mapping to start on Google.
Dynamic e-commerce prices move away from static prices (where items have a fixed cost) and
instead use data to set the price of items. Therefore, the more data that can be analyzed, the better the
optimal prices will be displayed. This means that the price no longer depends only on how much stock there
is. It uses a series of advanced analyzes about the product, the buyer, and the competitors to determine the
best price at a given time.
The strategy itself cannot completely eliminate uncertainty about consumer responses to different
prices, although there are many advantages to setting dynamic prices. In practice, addressing this issue
requires the efficient use of sales data. Sales data are data in time series that contain prices, along with those
sales and other features that could be useful in boosting sales (such as inventory, product seasonality, listing,
etc.). It is an input that a company should provide to build a dynamic pricing engine.
Because it is difficult to measure the performance of agents in real life, in this article we present an
artificial example, along with theoretically optimal prices for comparison. There are many different
approaches for training agencies to choose the optimal price for a certain period. We begin the process of
building a dynamic pricing system by building a sales prediction model that takes market status as input and
produces estimated sales. This model is trained on sales data by exploiting time series analysis and machine
learning methods. An important note here is that the price must be in a set of input features for a prediction
model. In addition, the characteristics should be of sufficient importance.
There are several possible approaches to defining a pricing strategy when using a forecast model:
• Optimization approach. This is a simple one-step optimization. Using the forecast model, you can set
different prices and see how they will affect sales and revenue. This is an optimization problem with a

116
 single variable over a continuous interval, which is defined by a minimum and maximum price. A typical
solution is to discretize a continuous range into a countable number of possible prices and choose the
price that leads to the highest income or another goal. An additional advantage of this method is that it
naturally avoids the problem of falling to a local minimum.
The disadvantage of this approach is that it only considers the income from a current time stamp, rather
than the cumulative income over a longer period, which is a better choice for maximization.
• Reinforcement learning (RL) is an alternative approach and one that solves the problem specified above. It
considers the associativity between the stages of setting succulent prices, thus being greedy in terms of
current time-marked revenues, but greedy in terms of cumulative income. For example, it might be
beneficial to set a lower price in the current time frame, lowering the potential income to later reach a
better market condition.

4. Reinforcement learning
Reinforcement learning, or RL, is one of the three main areas of machine learning, alongside
supervised learning, and unsupervised learning. Its main task is to develop algorithms that allow agents to
maximize their accumulated rewards while operating in an unknown environment. RL learns what to do
(politics) and how to map situations into action.
The idea of learning through interaction with an environment is very natural and, in some respects,
biologically inspired. As human beings, we continually interact with the world around us, trying to maximize
our reward (satisfaction, recognition, money, etc.).

Figure 3. How RL works (source [10])

In Reinforcement Learning (RL), the agent is rewarded for the right moves and punished for the
wrong ones. By doing this, the agent tries to minimize the wrong movements and maximize the right ones.
In setting dynamic prices, an agent should set optimal prices based on market conditions. In terms of RL
concepts, stocks are all possible prices and conditions, market conditions, except for the current price of the
product or service.
In fact, it is quite difficult to train an agent in an interaction with a real-world market. The reason is
that an agent should get a lot of samples from an environment, which means that it will take a long time as it
is a lengthy process. An agent should analyze a representative subset of the entire data set, trying different
actions. As a result, the company could lose a lot of money while training the agent, which would work sub
optimally.
An alternative approach is to use an environmental simulation. Using a forecast model, the reward
(e.g., income) can be calculated based on market conditions, and the stock is the current price. So, transitions
between states should only be modeled. The main disadvantage of the RL approach is that it is extremely
difficult to accurately simulate a market.
For instance, in the e-commerce dynamic pricing problem, we could map these concepts to:
• environment: marketplace
• state: lowest price in the market, day of the week, inventory levels, shipping values to a location, and
many others.
• agent: dynamic pricing algorithm
• action: to increase or to lower prices
• reward: total profit generated by the agent’s decisions

117
 Figure 4. E-commerce dynamic pricing problem map

To implement a dynamic pricing strategy in an e-commerce platform, an e-business owner must:

 Collect market data: The first step in implementing a dynamic pricing strategy is to collect market data,
including competitor prices, customer reviews of similar products, and buying trends over time.
 Choose the right dynamic price module: There are different dynamic price modules for different product
categories. Some modules are based on the type of data available and customer perceptions and
expectations.
 Choose the right type of dynamic pricing: There are also different types of dynamic pricing, which differ
depending on the season, competitors, stock availability, etc. Companies can choose to apply one or more
types of dynamic pricing depending on their needs.
 Apply dynamic price algorithms to collected data: A dynamic price algorithm analyzes the collected data
and offers prices or price ranges for different products accordingly. Old dynamic pricing algorithms are
based on classical statistical methods, such as Bayesian strategies, while newer algorithms use ML
methods, such as deep learning or consolidation learning for more accurate results.
 Update data: Prices change frequently, so price data needs to be collected frequently. Companies can use
web crawlers to automate the extraction of price data to get real-time data.

Advantages of reinforcement learning

 It can solve complex and higher order problems, the solutions obtained will be very precise, being very
similar to the human learning technique.
 This model will go through a rigorous training process that can take time, which can help correct any
errors.
 Due to its learning ability, it can be used with neural networks. This can be called deep consolidation
learning.
 Because the model is constantly learning, a mistake made earlier would be unlikely to occur in the future.
 Various problem-solving models can be built using reinforcement learning.
 The best part is that even when there is no training data, you will learn from the experience you have
from processing the training data.
 For various problems, which may seem complex, it offers the perfect models to solve them.

Conclusions
E-business owners, in today's competitive environment, face the challenge of adapting to market
changes, as lagging can mean losing leadership and revenue. Applying consolidation learning to dynamic
pricing can become an approach that brings an important change in the hierarchy of competitors, to be a
game changer for retail. Dynamic pricing can help retail business owners stay ahead of the market, and
consolidation learning can overcome dynamic pricing challenges. Amazon is without a doubt one of the
largest e-commerce platforms in the world, with over 300,000,000 active users. Amazon changes its prices
more than 2.5 million times a day to set prices lower than competitors. Since 2016, dynamic prices have
helped Amazon increase profits by 25%.

References
1. Reinforcement learning for fair dynamic pricing. Roberto Maestre, Juan Duque, Alberto Rubio, and Juan Arévalo - arXiv preprint
arXiv:1803.09967, 2018.
2. Dynamic pricing and energy consumption scheduling with reinforcement learning - Byung-Gook Kim, Yu Zhang, Mihaela Van Der
Schaar, and Jang-Won Lee. IEEE Transactions on Smart Grid, 2016
3. Dynamic pricing on e-commerce platform with deep reinforcement learning: a field experiment - Jiaxi Liu, Yidong Zhang
https://arxiv.org/pdf/1912.02572.pdf
4. Data-Driven Automated Dynamic Pricing for E-commerce, Jiří Moravčík https://dspace.cvut.cz/bitstream/handle/10467/94420/F3-
BP-2021-Moravcik-Jiri-data_driven_automated_dynamic_pricing_for_e_commerce.pdf

118
 5. Reinforcement Learning for Fair Dynamic Pricing Roberto Maestre, Juan Duque, Alberto Rubio, Juan Arevalo Intelligent Systems
Conference 2018 https://arxiv.org/pdf/1803.09967.pdf
6. https://techvidvan.com/tutorials/reinforcement-learning/
7. https://www.techradar.com/news/google-cloud-is-launching-an-ai-recommendation-engine-for-online-retailers
8. https://www.oecd.org/coronavirus/policy-responses/e-commerce-in-the-time-of-covid-19-3a2b78e8/
9. https://www.namogoo.com/resources/report/ecommerce-promotion-insights-survey-2021-shoppers-edition/
10. https://en.wikipedia.org/wiki/Reinforcement_learning

119

View publication stats