How to Survive Dynamic Pricing
Competition in E-commerce
Rainer Schlosser Martin Boissier Andre Schober Matthias Uflacker
Hasso Plattner Institute Hasso Plattner Institute adanbo GmbH Hasso Plattner Institute
Potsdam, Germany Potsdam, Germany Berlin, Germany Potsdam, Germany
rainer.schlosser@hpi.de martin.boissier@hpi.de andre.schober@adanbo.de matthias.uflacker@hpi.de
ABSTRACT sible from a given number of books (inventory level) in a
Pricing on e-commerce platforms is highly challenging. Sell- reasonable amount of time.
ers typically i) rival against dozens of competitors, ii) decide The pricing strategy of our project partner is character-
on prices for thousands of products, and iii) face steadily ized by a rule-based system that has been developed over
changing market situations. With respect to pricing, the the past years by carefully adjusting rules to lessons learned
challenge is to circumvent the curse of dimensionality to dy- from selling books on Amazon. As our project partner has
namically price products for a given market situation in a more than 10 years of experience in the market, we consider
timely manner. In this project, we create a stochastic pric- his strategy to be effective and accurate. However, market
ing model by analyzing recorded market data. This pricing dynamics are increasingly sophisticated making rule-based
model can be applied ad-hoc in less than a millisecond per strategies increasingly hard to handle and maintain.
item, allowing us to react immediately to new market situ- Our goal is to develop a pricing strategy that maximizes
ations. Our pricing approach is currently being applied in expected discounted long-term profits while taking into ac-
practice by a large German book seller on Amazon and out- count the constraints mentioned above. We seek to compute
performs the previous rule-based strategy by over 20% with data-driven pricing strategies that are applicable even for
respect to cash-in per book. large inventories.
2. DATA-DRIVEN PRICING MODEL
CCS Concepts
The project is devoted to revenue management [3] and
•Applied computing → Online shopping; E-commerce combines theory of dynamic pricing research and its practi-
infrastructure; Decision analysis; cal application [1, 2]. To be able to set up a dynamic model
in order to compute optimized prices, we need to estimate
Keywords sales probabilities. We use logistic regression analyses to
quantify how offer prices and specific market situations af-
Dynamic Pricing; Oligopoly Competition; Online Markets; fect sales. We consider up to 10 offer dimensions (e.g., price,
Demand Estimation quality, ratings, feedback count, shipping time) per competi-
tor for a particular market situation.
1. CHALLENGE
Modern market platforms such as Amazon Marketplace << table >>
or eBay are highly dynamic as sellers can observe the cur- OFFER
rent market situation at any time and adjust their prices isbn10 varchar
instantly. For sellers that handle large inventories, this dy- item_condition varchar
item_subcondition varchar
namic is hard to manage as an optimal pricing decision re- shipping_time varchar
quires handling a multitude of dimensions for each competi- feedback_rating varchar
tor (e.g., price, quality, shipping time, shipping costs, rat- …
offer_date datetime
ing). Moreover, financial aspects such as discounting as well
*
as inventory holding costs have to be taken into account.
1..*
In this project, we partner with adanbo GmbH. adanbo
<< table >> << table >> << table >>
is among the top 10 sellers for used books on Amazon in
ORDER STOCK PRICE
Germany with an inventory of over 80,000 distinct books
sku varchar sku varchar sku varchar
(ISBN), each with multiple items (1-20). Our seller can de- sold_date datetime isbn10 varchar price decimal
cide – to some extent – on the replenishment of used books price decimal in_date datetime upd_date datetime
(by choosing purchase prices). However, supply is limited out_date datetime
0..1 1 quality int 1 1..*
and it is not possible to directly reorder specific books.
Hence, the challenge is to extract as much profit as pos-
Figure 1: Overview of normalized data set.
Copyright is held by the author(s).
The data set that we use for the regression analysis con-
Development of a Data-Driven Dynamic Pricing Model Application of Dynamic Pricing Strategies
>20 M market situations
Estimate sales
Recorded
Our per month
Market probabilities for Efficient calculation
Offers
Situations specific market Stochastic Observe current of optimized price Adjust price on
Our
Realized
situations using Dynamic Pricing market situation for current market market place
Sales regressions Model situation
with customized
On-the-fly aggregation on streaming time- features
series data and feature construction
Self-correction via feedback loop of new observations (based on the applied strategy).
Figure 2: Two-phase process of the data-driven pricing strategy.
tains both the requested market situations from Amazon as
well as adanbo’s own data (offers, sales, and inventory; see Table 1: Comparison of adanbo’s and our data-
diagram in Fig. 1). Adanbo requests market situations for driven strategy.
Group size Sold Avg. price per sale
each offered book every two hours (i.e., >20 M market situ- adanbo’s strategy 3,535 22% 4.34 EUR
ations per month which result in >140 M single competitor data-driven strategy 5,502 17% 5.30 EUR
observations per month). We join this data on the fly with
adanbo’s price updates, placed orders, and stock data to
create the required observations and the corresponding fea- 3. RESULTS
tures. Working directly on the raw time-series data provides Our data-driven approach is currently applied by our pro-
us with more flexibility, e.g., when regressing only a sub- ject partner adanbo. We compare our strategy with adanbo’s
set of comparable market situations. We use 30 customized established rule-based strategy for two similar test groups of
features, e.g., the price rank of our offer price within the books (see Table 1). The data-driven strategy sells less ag-
competitors’ prices. The dependent variable is the number gressive and more profitable. Fig. 3 shows the ratio of the
of realized sales of a certain book in a certain time interval. average prices per sale over time. Around two weeks af-
As a result, we are able to predict sales probabilities for any ter the begin of the comparison, the advantage of the data-
offer price and for any market situation. driven strategy averages around a cash-in increase per book
Based on estimated (conditional) sales probabilities, we by approximately 20 percent.
set up and calibrated a suitable dynamic model. Using effi- Note, the model’s discount factor allows to control the
cient solution techniques, we are able to compute optimized strategy’s aggressiveness and in turn the speed of sales. As
prices for current market situations. The application of our a next step, we’d like to evaluate different levels of aggres-
dynamic pricing strategy works as follows: First, we observe siveness and their impact on profitability.
current market situations for our products, we then calcu-
late optimized prices according to the model, and finally
adjust prices on the market platform (see right-hand side of 4. CONCLUSION
Fig. 2). This procedure is repeated every two hours or in We presented a data-driven pricing approach for compet-
case of changing market situations. This way our strategy is itive sales applications. With our strategy applied, prof-
able to respond immediately to new situations as prices can its can be significantly increased. Moreover, by using the
be adjusted in milliseconds. Moreover, the new incoming model’s discount factor as a management instrument the
sales observations are used to further improve the strategy seller is able to smoothly balance profits, revenues, and the
by estimating demandofmore
Comparison accurately,
avg. price seetime:
per sale over Fig. 2. speed of sales.
Data-Driven vs. Adanbo's Strategy
1.8
5. REFERENCES
€ avg. price per sale (data-driven)
€ avg. price per sale (Adanbo)
1.6 [1] M. Chen and Z.-L. Chen. Recent developments in
1.4 dynamic pricing research: Multiple products,
competition, and limited demand information.
1.2
cash-in increase Production and Operations Management,
approx. 20%
1.0 24(5):704–731, 2015.
0.8 [2] B. D. Chung, J. Li, T. Yao, C. Kwon, and T. L. Friesz.
0.6
Demand learning and dynamic pricing under
27-M
ay-2
1-Ju
n-20
6-Ju
n-20
11-J
un-2
16-J
un-2
21-J
un-2
26-J
un-2
competition in a state-space framework. IEEE Trans.
016 16 16 016 016 016 016
Engineering Management, 59(2):240–249, 2012.
[3] K. Talluri and G. Van Ryzin. The Theory and Practice
Figure 3: Comparison of average price per sale over of Revenue Management. International series in
time: data-driven vs. adanbo’s strategy. operations research & management science. Kluwer
Academic Publishers, 2004.