Pricing on e-commerce platforms is highly challenging. Sellers typically i) rival against dozens of competitors, ii) decide on prices for thousands of products, and iii) face steadily changing market situations. With respect to pricing, the challenge is to circumvent the curse of dimensionality to dynamically price products for a given market situation in a timely manner. In this project, we create a stochastic pricing model by analyzing recorded market data. This pricing model can be applied ad-hoc in less than a millisecond per item, allowing us to react immediately to new market situations. Our pricing approach is currently being applied in practice by a large German book seller on Amazon and outperforms the previous rule-based strategy by over 20% with respect to cash-in per book.
Modern market platforms such as Amazon Marketplace or eBay are highly dynamic as sellers can observe the current market situation at any time and adjust their prices instantly. For sellers that handle large inventories, this dynamic is hard to manage as an optimal pricing decision requires handling a multitude of dimensions for each competitor (e.g., price, quality, shipping time, shipping costs, rating). Moreover, nancial aspects such as discounting as well as inventory holding costs have to be taken into account.
In this project, we partner with adanbo GmbH. adanbo is among the top 10 sellers for used books on Amazon in Germany with an inventory of over 80,000 distinct books (ISBN), each with multiple items (1-20). Our seller can decide { to some extent { on the replenishment of used books (by choosing purchase prices). However, supply is limited and it is not possible to directly reorder speci c books. Hence, the challenge is to extract as much pro t as posCopyright is held by the author(s). sible from a given number of books (inventory level) in a reasonable amount of time.
The pricing strategy of our project partner is characterized by a rule-based system that has been developed over the past years by carefully adjusting rules to lessons learned from selling books on Amazon. As our project partner has more than 10 years of experience in the market, we consider his strategy to be e ective and accurate. However, market dynamics are increasingly sophisticated making rule-based strategies increasingly hard to handle and maintain.
Our goal is to develop a pricing strategy that maximizes expected discounted long-term pro ts while taking into account the constraints mentioned above. We seek to compute data-driven pricing strategies that are applicable even for large inventories. 2.
The project is devoted to revenue management [
<< table >>
ORDER sku varchar sold_date datetime price decimal 0..1 << table >>
OFFER isbn10 varchar item_condition varchar item_subcondition varchar shipping_time varchar feedback_rating varchar … offer_date datetime
* 1..* << table >>
STOCK sku varchar isbn10 varchar in_date datetime out_date datetime 1 quality int << table >>
PRICE sku varchar price decimal upd_date datetime 1 1..* The data set that we use for the regression analysis conDevelopment of a Data-Driven Dynamic Pricing Model Application of Dynamic Pricing Strategies Our Of ers
Self-correction via feedback loop of new observations (based on the applied strategy). tains both the requested market situations from Amazon as well as adanbo's own data (o ers, sales, and inventory; see diagram in Fig. 1). Adanbo requests market situations for each o ered book every two hours (i.e., >20 M market situations per month which result in >140 M single competitor observations per month). We join this data on the y with adanbo's price updates, placed orders, and stock data to create the required observations and the corresponding features. Working directly on the raw time-series data provides us with more exibility, e.g., when regressing only a subset of comparable market situations. We use 30 customized features, e.g., the price rank of our o er price within the competitors' prices. The dependent variable is the number of realized sales of a certain book in a certain time interval. As a result, we are able to predict sales probabilities for any o er price and for any market situation.
Based on estimated (conditional) sales probabilities, we set up and calibrated a suitable dynamic model. Using e cient solution techniques, we are able to compute optimized prices for current market situations. The application of our dynamic pricing strategy works as follows: First, we observe current market situations for our products, we then calculate optimized prices according to the model, and nally adjust prices on the market platform (see right-hand side of Fig. 2). This procedure is repeated every two hours or in case of changing market situations. This way our strategy is able to respond immediately to new situations as prices can be adjusted in milliseconds. Moreover, the new incoming sales observations are used to further improve the strategy by estimatinCgomdepmariasonndomfavogr.epraiccecpuerrastaelelyo,vesreteimFei:g. 2.
Data-Driven vs. Adanbo's Strategy ) 1.8 irven )nob1.6 -d a taad (dA1.4 (e la
e lrsa rse1.2 e p icep irce1.0 rp .gp .vag va€0.8 € 0.6 cash-in increase approx. 20%
adanbo's strategy data-driven strategy
Our data-driven approach is currently applied by our project partner adanbo. We compare our strategy with adanbo's established rule-based strategy for two similar test groups of books (see Table 1). The data-driven strategy sells less aggressive and more pro table. Fig. 3 shows the ratio of the average prices per sale over time. Around two weeks after the begin of the comparison, the advantage of the datadriven strategy averages around a cash-in increase per book by approximately 20 percent.
Note, the model's discount factor allows to control the strategy's aggressiveness and in turn the speed of sales. As a next step, we'd like to evaluate di erent levels of aggressiveness and their impact on pro tability.
We presented a data-driven pricing approach for competitive sales applications. With our strategy applied, profits can be signi cantly increased. Moreover, by using the model's discount factor as a management instrument the seller is able to smoothly balance pro ts, revenues, and the speed of sales.