Point of Sale (POS) Data from a Supermarket: Transactions and Cashier Operations

Abstract

As queues in supermarkets seem to be inevitable, researchers try to find solutions that can improve and speed up the checkout process. This, however, requires access to real-world data for developing and validating models. With this objective in mind, we have prepared and made publicly available high-frequency datasets containing nearly six weeks of actual transactions and cashier operations from a grocery supermarket belonging to one of the major European retail chains. This dataset can provide insights on how the intensity and duration of checkout operations changes throughout the day and week.

Dataset: Supplementary data to this article.

Dataset License: CC-BY-NC

1. Summary

Retail store operations are an active and a relatively wide area of research. In a recent study, Mou et al. [1] reviewed 255 publications from 32 operations research, retailing and management journals over the period 2008–2016 and categorized works by distinguishing seven operational decisions pertinent to store management. These included: (1) demand forecasting; (2) in store logistics; (3) inventory management; (4) assortment and display; (5) product promotion; (6) checkout operations and (7) employee management. Interestingly, the authors argue that in particular checkout operations will attract more attention in the near future.

As of today, however, only a few studies related to checkout operations have been published [2,3,4,5]. The likely reason is the (un)availability of recent and representative point of sale (POS) data. Even if such data is analyzed, it is often “proprietary and therefore not available to researchers at large”, as in the case of the Mas and Moretti dataset [6]. With this in mind, we have prepared and made publicly available high-frequency datasets containing nearly six weeks of actual transactions and cashier operations from a grocery supermarket belonging to one of the major European retail chains. This dataset can provide insights on how the intensity of checkout operations changes throughout the day and throughout the week. Hence, it can be used as a starting point for building realistic agent-based or forecasting models of customer behavior. In practice, such data—if available in real-time—can augment detectors or video content analysis technologies (VCA) used to count customers inside a store [7] and yield better predictions of the demand for opened checkouts. On the other hand, it can be used to provide feedback, e.g., in the form of voice or visual messages, about the current or near-future state of the checkout zone, with the ultimate objective of speeding up the checkout process and increasing consumer satisfaction [8].

2. Data Description

The data was retrieved from checkout/POS system logs stored in XML files, which contained various low-level transactional data. Once extracted, it was aggregated into six CSV files with the most important information about (i) transactions; and (ii) cashier operations, see

Table 1. Transactions data (POS_transactions_*.csv files): fields, data types and descriptions.

Field	Type	Description
WorkstationGroupID	Integer	Type of checkout: 1—service, 8—self-service
TranID	Numeric	Transaction ID (date, store ID, checkout ID, sequence no.)
BeginDateTime	Date/Time	Date and time of transaction start
EndDateTime	Date/Time	Date and time of transaction end
OperatorID	Integer	Unique cashier ID
TranTime	Integer	Transaction time in seconds 1
BreakTime	Integer	Break (including idle) time in seconds 2
ArtNum	Integer	Number of items, i.e., basket size
TNcash	True/False	Cash payment flag (true when transaction paid in cash)
TNcard	True/False	Card payment flag (true when transaction paid by a card)
Amount	Numeric	Transaction value

¹ Computed for the n-th transaction as: TranTime(n) = EndDateTime(n) − BeginDateTime(n). ² Computed for the n-th transaction as: BreakTime(n) = BeginDateTime(n) − EndDateTime(n − 1), i.e., the latter is from the previous transaction.

and

Table 2. Cashier operations data (POS_operator_logs_*.csv files): fields, data types and descriptions.

Field	Type	Description
WorkstationGroupID	Integer	Type of checkout: 1—service, 8—self-service
WorkstationID	Integer	Unique checkout ID
TranID	Numeric	Transaction ID (date, store ID, checkout ID, sequence no.)
BeginDateTime	Date/Time	Date and time of transaction start
OperatorID	Integer	Unique cashier ID
Items	Text	Operation identifier 1

¹ Admissible values: OperatorSignOn—cashier log-in, OperatorSignOff—cashier log-off, OperatorLock—start of cashier’s break, OperatorUnLock—end of cashier’s break.

, respectively. The data concerns retail operations in a grocery supermarket located in a large city in Southern Poland, equipped with manned (service) and self-service checkouts. The checkout zone is composed of a single waiting line for each service checkout and one waiting line for all self-service checkouts. The data covers three nearly two-week periods: (i) 7 to 19 December 2017; (ii) 13 to 26 February 2019; and (iii) 28 March to 10 April 2019. Please note that the new regulations introduced in Poland in 2018 banned shopping on some Sundays, generally two Sundays per month in 2018 and three Sundays per month in 2019, disrupting the rather regular 7-day pattern observed until the end of 2017. Supermarkets have reacted by extending the working hours on Fridays and Saturdays, while customers had to adapt to the changes in opening hours. The two pairs of datasets from 2019 include one working (24 February, 31 March) and one non-working Sunday (17 February, 7 April) each.

3. Methods

The two datasets were extracted from checkout/POS system log files of a supermarket. The logs are archived in XML files and contain various low-level transactional data, most of which is not relevant for the analysis of transactions or cashier operations. A small fragment of a sample log file is depicted in Figure 1. Note that the checkout service generally consists of three separate activities: scanning (registration) of articles, payment and bagging (including idle time). POS logs include the exact times of starting (registration of the first article in the basket; BeginDateTime) and end times of the transactions (EndDateTime).

However, the data has its limitations. For instance, the registered end time is not exactly the time when the payment is made and the operation is terminated. In particular, for cash payments EndDateTime does not cover the activity of giving back the change to the customer, while the time between transactions (BreakTime) retrieved from POS data includes the idle time between two operations, which actually is not part of the service activity. However, given that idle times are very rare during peak hours, by analyzing only periods of high activity (particularly Thursdays 10 a.m. to 1 p.m., Fridays and Saturdays 11 a.m. to 2 p.m.) we can essentially eliminate the impact of idle times and obtain information about the service time itself. The timeline of the checkout service (scanning, payment and bagging) and the times retrieved from POS logs are illustrated in Figure 2.

Regarding queue management/modeling, the data does not contain customer arrival information. However, it is possible to extract an approximate arrival rate. For instance, one can combine a theoretical model (e.g., a Non-Homogeneous Poisson Process, NHPP [9]) with transactional data, i.e., approximate the arrival rate of a NHPP at a certain hour by the average number of transactions in a time window (e.g., +/− 30 min) around this hour. Such an approach would yield an edge over completely theoretical arrival process models typically used in publications concerning modeling queues in supermarkets. Finally, despite the fact that balking and reneging unarguably take place, our own observations and interviews with line workers suggest that they are so incidental, that they do not affect significantly the queuing process.