Live shopping promotions: which categories should a retailer discount to shoppers already in the store?

While previous research has mostly focused on share-of-wallet—using a variety of similar but distinct definitions (e.g., Chen and Steckel 2012; Du et al. 2007; Glady and Croux 2009)—our study involves the related construct share-of-transactions. In the following section, we briefly define both terms and provide an overview of existing research, which we summarize in Table 1.

Most commonly, the total expenditures of a customer in a particular product category define his/her size-of-wallet, while the proportion of category-level expenditures at a focal company defines his/her share-of-wallet (e.g., Kumar and Reinartz 2012; Jang et al. 2016). We, however, focus on the proportion of a customer’s category-level transactions at a retailer. This is equivalent to share-of-visits (cf. Mägi 2003) but at the category instead of basket level. We do so, as we are particularly interested in purchase incidences and their timing. Moreover, our share-of-transactions metric could be extended to share-of-wallet when it is multiplied by a customer’s category-level spending. We model share-of-transactions as the long-term probability that a customer purchases at a focal retailer in a particular category. In addition, we focus on customers’ outside potential (cf. potential-of-wallet, Glady and Croux 2009) involving their transactions at competing retailers. These are precisely the purchases that a focal retailer would like to turn into his/her own. To the best of our knowledge, we are the first study to analyze category-level share-of-transactions in grocery retailing based on the transaction data of a single retailer. So far, studies in grocery retailing have relied on customers’ self-reports (Mägi 2003) or panel data (Meyer-Waarden 2007) and have analyzed share-of-wallet measures on store level (Mägi 2003).

Existing studies that have focused on multiple product categories in financial services have directly incorporated cross-category effects (symmetric effects for ten categories: Du et al. (2007); asymmetric effects for two categories: Jang et al. (2016)). However, the variety of product categories in grocery retailing complicates the inclusion of cross-category effects. For this reason, we have decided not to include separate cross-category effects but to include the impact of basket-level purchasing behavior on category-level share-of-transactions via semi-informed priors in the Bayesian estimation process. This approach incorporates cross-category purchase behavior through basket information while decreasing complexity.

Our approach to derive category-level share-of-transaction estimates requires only transaction data and does not rely on data augmentation strategies through customer surveys. In line with Glady and Croux (2009), who were the first to use customer survey data only for model validation (not for model estimation), we use panel data to validate our models, which we ultimately run on the data from a single grocery retail chain.

The key objective of share-of-wallet research is to identify and target customers with outside potential (i.e., high potential-of-wallet) (Chen and Steckel 2012; Jang et al. 2016; Du et al. 2007; Glady and Croux 2009). We follow this line of research as we select categories in which customers still have outside potential (i.e., low share-of-transactions, high potential-of-transactions) and we extend this research by incorporating information on true category-level interpurchase times in the category selection process.

As in grocery retailing, personalized price promotions occur frequently, we control for their impact on customers’ share-of-transactions. To do so, we estimate a measure of customers’ reaction to purchases on personalized price promotions; i.e., an increase in share-of-transactions.

Our research contributes to the existing research on (personalized) price promotions—such as coupons—and the crucial role of timing. Coupons are still widely used by customers, as a recent survey by Inmar (2018) highlights: 88% of customers use coupons, and 83% of the customers using coupons indicate that the coupons influence their shopping behavior. At the same time, demand for personalization has increased; according to a study by Capgemini (2017), 65% of customers become frustrated when promotions are not personalized.

In our empirical application, we focus on categories in which customers have made purchases but are still able to expand their share-of-transactions because they frequently purchase at competing retail chains (cf. Gedenk et al. 2010; Osuna et al. 2016; Venkatesan and Farris 2012). This enables us to leverage the category-level share-of-transactions information that we infer based on customers’ transaction data at a retailer. Our research follows Osuna et al. (2016), who derive brand-level and category-level guidance on drivers of redemption rates and incremental sales (cf. p. 248, Table 5). We extend their perspective by incorporating (coupon) timing when selecting the categories for which to offer personalized price promotions at a particular point in time.

There has also been research on the timing of in-store coupons, illustrating that the average observed category-level interpurchase time does not coincide with the point in time where coupon redemption is most likely (Vuckovac et al. 2016). In contrast with our study, however, Vuckovac et al. (2016) do not relate coupon redemptions to the true category-level interpurchase times (i.e., involving observed and unobserved purchases) nor do they model potential effects of promotions on share-of-transactions (i.e., endogeneity).

Regarding the optimal timing of personalized promotions, Zhang and Krishnamurthi (2004) are among the first to acknowledge its importance, along with discount depth, by investigating what discount to offer and to whom on a given shopping trip. They optimize discounts conditional on store visit while accounting for the importance of timing through inertia and variety-seeking as time-varying behavioral traits. Johnson et al. (2013) take the analysis of customized promotions one step further by allowing customized promotions to influence customers’ purchase timing. They analyze what discount to offer to whom and at what point in time from the perspective of a manufacturer. For their analyses, they use panel data and the data of a single retail chain which they treat similar to panel data (i.e., no store switching). We provide a summary of the relevant literature on the timing of personalized price promotions in Table 1.

Loyalty program datasets may contain hundreds of thousands or millions of customers with dozens of shopping trips. Therefore, considering cross-category or even cross-product effects in models for personalized targeting is typically not feasible when traditional modeling approaches are used. However, important recent contributions (see Table 1) in the field of machine learning (ML) even allow for considering cross-product effects in a sparse form. ML approaches for product choice modeling show an exceptional performance and good scaling properties for targeted price promotions (Gabel and Timoshenko 2021; Jacobs et al. 2016; Ruiz et al. 2020).

Ruiz et al. (2020), for instance, model how customers select products and consider interactions between products and price changes. They test their efficient posterior inference algorithm on a large dataset containing shopping baskets from a grocery chain and identify complementary and substitutable product relationships.

Recently, Gabel and Timoshenko (2021) developed a deep-learning model at the product level to optimize the assignment of personalized in-store coupons. They find that their model performs particularly well when cross-category promotional effects are pronounced. Their highly flexible nonparametric product-level model allows for the detection of a broad variety of interesting relationships in an exploratory way. One of their key findings (Gabel and Timoshenko 2021, p. 18) is that the last purchase of a product in a category is extremely important in improving their forecasts compared to other benchmarks. More specifically, the more recent the last purchase, the lower the likelihood of a purchase in that category in the near future. Their finding is well aligned with our more restrictive parametric assumptions and (like our model) in full contrast with RFM-based targeting approaches that many large retailers use (assuming that the response probability is higher when a customer’s last purchase is more recent). Thus, our paper complements the more explorative nonparametric work of Gabel and Timoshenko (2021) with a parametric model. Differences between their product-level approach and our category-level model can also be found in terms of customization effort and generalizability of the findings. While Gabel and Timoshenko’s (2021) model is very convenient (no customization effort) in the sense that no categories or cross-product effects have to be laboriously defined, an advantage of our proposed methodology is that our parametric model yields insights that (due to the parametric nature) can easily be generalized for use in heuristic approaches.

In contrast with the work by Johnson et al. (2013) or Gabel and Timoshenko (2021), our study focuses on the category level instead of the brand or product level, taking on the perspective of a grocery retail chain instead of a single manufacturer (e.g., Ailawadi et al. 2006). Both, product-level models and category-level models have their merits. If promotions are offered at the category level, they provide more freedom of choice to customers than promotions that are limited to specific brands (e.g., discounting all shampoo brands vs. discounting a specific brand). Thus, if customer satisfaction is a retailer’s key objective, category-level promotions seem more attractive as they preserve customers’ freedom of brand choice. In this context, the private label share in the assortment may also influence whether a category- or product-level approach to promotions should be pursued. Retailers, like Aldi or Migros that have a large private label share may prefer a category-level approach in order to let the consumer make the final brand or product choice within the category. In contrast, retailers with many brands in their assortment like EDEKA or LIDL may leverage the competition between manufacturers and let manufacturers pay for “their” promotions. In such a case the product-level approach of Gabel and Timoshenko (2021), which provides information about substitutes (e.g., the likelihood of buying a Coke decreases with a recent purchase of a Pepsi), is more closely linked to manufacturer’s objectives. However, retailers do not only focus on the costs of promotions. As mentioned above, an important objective for retailers is to increase their customers’ category-level share-of-wallet. In this case, the category-level perspective seems particularly attractive since it is directly linked to this important objective. An advantage of our approach is that we model purchases at the competition, which is necessary to estimate a focal retailer’s share-of-transactions.

With the retailer perspective in mind, we propose to target customers who are about to cover their category-specific needs at competitors of the focal retailer, as they have a comparatively low share-of-transactions. Our targeting approach is based on information about true category-level interpurchase times, customers’ category share-of-transactions, and customers’ sensitivity to promotions.

In particular, we suggest the following two-step in-store targeting approach to grocery retailers. At the beginning of every shopping trip, a customer scans his/her loyalty card and receives a set of (in our case seven) personalized price promotions on a printout. (1) The system used to determine the set of personalized price promotions per customer should use the true category-level interpurchase time to decide for which categories a promotion would be well timed at that particular point in time. (2) Among these categories, the system should select the ones where the particular customer has a relatively low share-of-transactions (i.e., highest potential-of-transactions). This ensures that customers get personalized price promotions that are relevant while the retailer uses the opportunity to turn competitors’ sales into own sales. Note, that this is in strong contrast to RFM-based approaches which prioritize recent and frequent buyers.

1. Individual true category-level interpurchase times \(T_{ir}^{\prime }\) (involving observed and unobserved purchases) are Erlang (2, β_i), β_i ≥ 0, distributed with \(t_{ir}^{\prime }\) describing the rth true interpurchase time of customer i.

The expected mean of the Erlang (2, β_i) distribution is 2/β_i and describes a customer’s average true category-level interpurchase time. The Erlang assumption identifies our model as it determines which purchases the focal retailer does and does not observe (Chen and Steckel 2012).³

2. A customer’s individual probability of purchasing in a given product category at the focal retailer (conditional on his/her category-level purchase) is determined by a Bernoulli process. The focal retailer observes a category-level purchase of customer i with probability p_i. We call this share-of-transactions (i.e., ShoT_i) and assume that the customer’s purchase probabilities across categories are in general independent and only related due to the basket prior of the ShoT_i.

3. When a customer purchases on promotion in a particular category at the focal retailer, this increases his/her category-level probability of purchasing at the focal retailer. As this positive effect decreases over time, a customer i’s share-of-transactions at point n in time is given by:

With δ_i increase in his/her category-level share-of-transactions (i.e., δ_i ≥ 0), tslpp_in time since last purchase on promotion at point n in time, a_i decay parameter (i.e., a_i ≥ 0).

4. We do not specify a dropout process but assume that the data are right-censored. Thus, customers remain active customers of the focal retailer during the complete observation period in the respective categories.

5. To account for the impact of basket-level share-of-transactions (i.e., store-level purchase behavior) on category-level share-of-transactions at the focal retailer, we use a semi-informed prior.

It follows from our assumptions that the expected observed interpurchase time T_ir of an individual i in a category at point n in time is:

While the estimation of true category-level interpurchase times and of the share-of-transactions both require information that is unknown to a single retailer, we can jointly estimate them using the transaction data of a single retailer (Chen and Steckel 2012).

Given customer i’s vector of parameter estimates at point n in time θ, we can derive the probability density function for observed category-level interpurchase times T_ir. For customer i, the distribution of observed interpurchase time T_ir in a category at point n in time is described by the following equation:

The term Erlang[t_ir|2(m + 1), β_i] describes the application of the additivity property of the Erlang-k distribution, i.e., the sum of m independent random variables following an Erlang (k, β_i) distribution is Erlang (km, β_i) distributed. As Eq. (4) provides the probability density function for the time between observed purchases in a product category, we have to allow for m unobserved purchases (at competitors) to lie between two observed purchases (at the focal retailer). The geometric distribution described by the first term gives the probability that m unobserved purchases lie between two observed purchases of customer i in a particular category at point n in time.

For every customer in every category, the analysis begins with his/her first observed purchase. When estimating the model, we set up the following likelihood function for observed interpurchase times T_ir, where i (to I) denotes the index for the customer and r (to R_i) denotes the rth observed interpurchase time indexing the point n in time as follows:

The second part of the likelihood function accounts for right censoring and describes the probability of observing no purchase between the last observed purchase and the end of the observation period (t_h) (Chen and Steckel 2012).

To derive model estimates, we follow a Bayesian sampling approach. To obtain individual priors, we run the model (in a reduced form without the impact of personalized price promotions) on customers’ basket-level purchases deriving an overall basket-level share-of-transactions for each customer, which is then used as the prior distribution for our category-level analysis. As sparse data and increasing complexity prevent us from including cross-category effects, the use of a semi-informed prior allows us to link store choice (i.e., basket-level information) to category-level purchases. We provide detailed information on the Bayesian estimation procedure in Appendix 2.

Purchasing on promotion is likely to interfere with a customer’s purchase routine and therefore needs to be incorporated in the model to avoid biased model estimates. Per se, our model structure offers two potential avenues for modeling the influence of purchases on promotion on customers’ purchase behavior: through customers’ probability of purchasing at the focal retailer in a particular category (conditional on category purchase, i.e., the share-of-transactions) and through the true category-level interpurchase times (i.e., via the rate parameter of the Erlang distribution). We pursue the former approach, as it implies that a purchase on promotion affects observed interpurchase times through its impact on share-of-transactions but does not affect the true category-level interpurchase time (cf. Equation (2)). This is in line with our proposition that promotions can be used to turn competitors’ sales into one’s own (i.e., an increase in customers’ share-of-transactions).

Figure 1 exemplifies how a customer’s share-of-transactions is influenced by purchases on promotion. Every purchase on promotion triggers an increase in a customer’s category-level share-of-transactions of the same magnitude (i.e., δ_i ≥ 0). This increase levels off over time (i.e., tslpp_in increases and starts again at zero with the next purchase on promotion) (Pauwels et al. 2002). To allow for heterogeneity in the impact of time on the increase in share-of-transactions, we include a customer- and category-specific decay parameter (i.e., a_i ≥ 0).⁴ Moreover, we normalize the decreasing impact of time since the last purchase on promotion, with a customer’s average interpurchase time in the respective product category (i.e., 2/β_i) to facilitate cross-category comparisons.

The increase in share-of-transactions in reaction to a purchase on promotion is conditional on a customer’s purchase on promotion. Thus, the expected effect of such a promotion on share-of-transactions is given by the category-specific probability to purchase on promotion times the average share-of-transactions increase effect.

For our model application, we use the transaction data from the loyalty program of a European grocery retail chain. Our data cover almost two years (i.e., 05/2015 until the end of 03/2017) and involve transactions of loyalty cardholders in 149 stores of the grocery retail chain in a major European country. For our analysis, we focus on a subset of loyalty cardholders with a sufficient number of transactions for fitting and evaluating the model (i.e., six purchases and at least one redemption of a personalized price promotion in the fourteen-month estimation set, and four purchases in the nine-month holdout data set). We select a subset of 3,335 loyalty cardholders and their transactions in seven product categories. The product categories are comparable to those analyzed in the panel data set but have been complemented by two fresh product categories (i.e., bananas and tomatoes, for which the grocery retail chain issues many personalized price promotions).

At the focal grocery retail chain, loyalty cardholders remain anonymous, as no personal information except for their transaction data is stored. The transaction data also contain information on personalized price promotions received and redeemed, starting at the point in time when the focal retailer initiated this loyalty program. Thus, our data include information on the first redemptions of all loyalty cardholders. To distribute the personalized price promotions among loyalty cardholders, the retailer uses an in-store coupon system, i.e., customers scan their loyalty card at the store entrance and receive a list of seven personalized price promotions that can be redeemed during their current shopping trip.

Such in-store coupons mitigate one potential drawback of customized checkout coupons: the temporal distance between receipt and redemption (Zhang and Wedel 2009). As coupon receipt coincides with coupon usage, such in-store coupons are expected to result in higher redemption rates than traditional checkout coupons as they work through pull (from the customer) instead of push (from the retailer). While customers cannot be lured to the store through personalized price promotions because they are received on-site, an in-store system can be useful to trigger unplanned category purchases (Briesch et al. 2009; Heilman et al. 2002; Inman et al. 2009).

Both the retailer and manufacturers can use the in-store coupon system to distribute coupons among the retailer’s loyalty cardholders. To choose the seven most promising personalized price promotions for loyalty cardholders, the proprietary system calculates the similarity between the loyalty cardholder’s transaction history and the active promotional campaigns, taking into account heterogeneity in price sensitivity when assigning the individual discount. Our analysis focuses on promotional timing, the customer’s share-of-transactions in the respective product category, and his/her increase in category-level share-of-transactions when redeeming a personalized price promotion. We neglect the discount height because it is already covered in the current assignment procedure. As we do not have information on regular one-size-fits all price promotions (for instance, via store flyers), we only capture the impact of redeemed personalized price promotions (i.e., increase in share-of-transactions). We provide summary statistics of the grocery retail data in Table 3.

The descriptive statistics show that there is considerable variation in the number of purchases, purchase spending, quantities, and observed interpurchase times across categories. Similarly, the promotion intensity (i.e., the number of received price promotions in the observation period ranges from 5.8 to 21.0) and redemption rates (ranging from 0.18 to 0.41) vary strongly across categories.

In addition to our model validation using panel data, we also validate our model on the data set provided by the grocery retail chain. As the data set only contains information on observed interpurchase times, we use those for model validation following an estimation-holdout approach. The results of this estimation-holdout validation are comparable to the results using panel data, for which we provide more detailed information in Appendix 3.

The focal retailer can leverage the insights from our modeling approach when assigning personalized price promotions via the in-store coupon system. In particular, every time a customer scans his/her loyalty card, the system should use the true category-level interpurchase times to determine for which categories a promotion would be well timed at that particular point in time. These should be the categories to be considered when distributing personalized price promotions to a customer. Then, the system should look at the customer’s share-of-transactions in the categories to be considered and select the categories with the lowest share-of-transactions. This is because at the focal retailer, customers with a low share-of-transactions tend to have a higher share-of-transactions increase in reaction to a purchase on promotion. We illustrate this relationship with a third-degree fitted polynomial in Fig. 2.

As the personalized price promotions a customer gets are limited to seven, the focal retailer should focus on categories that combine the best timing and highest transaction potential according to our proposed approach.

To evaluate ex post whether our proposed model indeed helps target customers with personalized price promotions that are most likely to satisfy their demands at competing retailers, we compare our approach with an alternative RFM-based targeting approach.

In practice, heuristics such as RFM are often used to determine which customers to address with targeted marketing campaigns (e.g., personalized price promotions) (Verhoef et al. 2002). Such heuristics often achieve similar results, as do more advanced modeling techniques, as Wübben and von Wangenheim (2008) note. The popularity of the RFM model and its variants among practitioners arises from their simplicity. Information on RFM is observable, easy to obtain (also based on an anonymous loyalty card), easy to understand, and easy to analyze. RFM methods can even be applied on the product and category level (Heldt et al. 2021) which is particularly useful when assigning personalized price promotions. Even though companies (such as retailers and providers of targeting solutions) often keep their targeting rules confidential, customers who receive targeted price promotions have been shown to have higher recency, frequency, and monetary value scores than those who do not receive targeted promotions (Park et al. 2018). This is because recency and frequency provide information on whether a category is relevant for a customer while monetary value provides insights on customer’s price sensitivity. Therefore, we use the RFM approach as a relevant industry benchmark. Besides RFM-based targeting, another frequently implemented method to generate personalized price promotions is based on a customer-product matrix and the respective cosine similarity between a decomposed customer- and a product-vector (cf. Levy and Goldberg 2014). Since all the personalized price promotions in our dataset were generated by a CS-based recommendation engine, this approach serves as a basic benchmark for the RFM-based targeting approach and for our proposed targeting method.

We apply the RFM-based approach and our live targeting method to the data of the focal grocery retail chain. To do so, we take the data period that we used to estimate our model (i.e., 05/2015 until the end of 06/2016) to determine two groups of target customers for personalized price promotions. Then, we evaluate the target groups based on redemption rates, the parameters of our proposed model, and RFM-scores.

For the RFM-based target group, we assign scores for recency, frequency, and monetary value (i.e., total category-level revenues) to all customers in the respective product category. In particular, we assign values ranging from 1 to 5 to the recency, frequency, and monetary value quintiles and sum these values to receive a final score per customer and product category, which ranges from 3 to 15. This corresponds to assigning equal weights to all elements of the RFM model. We then select the customers with the highest RFM scores.

To determine the target group based on our proposed model, we follow a two-step process. First, we consider all customers who receive personalized price promotions at a point in time that is in line with our proposed model. Second, we rank order these customers according to their share-of-transactions estimate and select the customers per product category with the lowest share-of-transactions. This is in contrast with the RFM approach, which prioritizes customers with high frequency and high monetary value (i.e., proxy for a high share-of-wallet) as well as recent category-level purchases.

The number of customers per category who received promotions whose timing is in line with our proposed model is 980. We consider the promotional timing of a personalized price promotion to be in line with our proposed model when the deviation from the point in time for which we expect the highest likelihood of redemption was lower than or equal to 20%.⁶ The categories in which the most customers received well-timed promotions are the fresh categories, namely, bananas, tomatoes, and bread (in total 870 of the 980 customers).

Since only 980 customers received well-timed promotions according to our proposed model, this serves as an upper limit for customer selection (i.e., share-of-transactions is not taken into account). To consider share-of-transactions, we then select the top 10% or the top 50% of customers with the lowest share-of-transactions (i.e., 98 or 490 customers). Importantly, we select customers who have received at least one promotion at the point in time suggested by our model during the application period; however, those customers could also have received and redeemed personalized price promotions whose timing was not in line with our proposed model.

When we apply the ex post selection strategies using the RFM-based approach, we select the customers with the highest RFM score to match the respective number of customers under our proposed approach (i.e., 98 or 490). The customers selected under the RFM-based approach have received at least one personalized promotion during the application period. Table 4 provides an overview of our ex post selection strategies.

We compare the three targeting approaches. The basic cosine similarity-based recommendations, the RFM-based approach and the one using the insights from our proposed model—and present the aggregated results in Table 5.

The results for the standard method of the retailer of generating targeted recommendations (i.e., CS) show a redemption rate of 27.7%. The results show that the redemption rate for RFM-based targeting (25.5% for top 10%, 29.7% for top 50%), is not significantly different from CS-based recommendations. The redemption rate in the target group based on live targeting (42.4% for top 10% and 33.8% for top 50%), in contrast, is significantly higher than the redemption rates in the CS- and RFM-based target groups. As expected, the increase in redemption rates for our proposed targeting method as compared to the other two approaches is higher when only 10% of the customers are targeted as compared to the situation when less stringent selection criteria (i.e., 50%) are applied. On average (over targeting 10% and 50% of all customers), the proposed targeting method leads to an increased redemption rate of 10.5 percentage points versus RFM-based targeting.

In addition to the redemption rates, Table 5 also indicates that the three targeting methods differ in terms of several characteristics. Since live targeting and CS-based targeting approaches do not select based on RFM scores, these scores are clearly below those of the RFM-based approach (14.03 for top 10% and 12.71 for top 50%). Naturally, the share-of-transactions parameter is lowest while the increase parameter is highest for the target groups based on our proposed live targeting model.

Our analyses reveal that most of the personalized price promotions are not well timed under CS- and RFM-based targeting approaches (as indicated by the lower redemption rates) which supports our focus on promotional timing. Next, we conduct further analyses on revenues and purchase frequency.

As a second illustration of the power of live targeting vs. RFM-based⁷ targeting, we use a matching analysis and measure effects on revenues and purchase frequencies. Using propensity-score matching, we estimate and compare the impact of targeting as opposed to not targeting customers under both approaches (e.g., Gensler et al. 2012; Ma 2016).

Following the RFM-based approach, a customer’s propensity to receive a promotion during the application period as opposed to no promotion is a function of his/her respective recency, frequency, and monetary value scores. Following our suggested targeting approach, a customer’s propensity to receive a (well-timed) promotion (as opposed to no promotion) is a function of his/her estimated true category-level interpurchase time, his/her share-of-transactions, and his/her response to purchases on promotion (i.e., share-of-transactions increase). We match customer pairs (one receiving a promotion under the respective targeting approach, the other one not) within a product category applying one-to-one nearest neighbor matching. As a distance measure, we use the Mahalanobis distance (cf. Gensler et al. 2012).

Propensity-score matching occurs based on customers’ transactions during the selection period and either takes the customers’ recency, frequency, and monetary value scores into account (RFM-based targeting approach) or the customers’ estimated true category-level interpurchase time, share-of-transactions, and share-of-transactions increase (our proposed targeting approach). We evaluate purchase frequency and revenues based on customers’ transactions during the application period. Table 6 contains the results of the revenue and purchase frequency analysis aggregated across product categories. Although fresh product categories (i.e., bananas, tomatoes, bread) drive our results as they contain the majority of observations (i.e., 74% under the RFM approach and 87% under our proposed approach), our results are stable across all categories. We find that both targeting approaches have a positive impact on customers’ purchase frequency and revenues. However, live targeting triggers greater absolute and relative increases in purchase frequency and revenues when compared to RFM-based targeting. For average purchase frequencies per customer and category, we find a significantly higher increase (p < 0.001) under our proposed targeting approach of 1.548 units (i.e., 113.74 percent) as opposed to 0.945 units (i.e., 69.54 percent) under the RFM-based targeting approach; i.e., an increase of 44.2 percentage points in terms of purchase frequency. The same holds for average revenues per customer and category as we find a significantly higher increase (p < 0.001) of 2.20€ (i.e., 99.28 percent) as opposed to 1.27€ (i.e., 56.96 percent) under our proposed targeting approach; i.e., an increase of 42.32 percentage points in terms of revenues. There are no significant differences between the control groups of both targeting approaches.

Thus, when selecting customers based on our model insights as opposed to selecting customers based on an RFM-based approach, the focal retailer can achieve higher purchase frequencies and revenues in addition to higher redemption rates. This also holds when we only select customers with particularly high RFM scores (i.e., ≥ 10) under the RFM approach and customers with rather low share-of-transactions parameters (i.e., ≤ 0.5) under our proposed approach.

When implementing live targeting, we recommend companies to first focus on customers for which there are enough category-level transactions to estimate our model. However, a retailer typically wants to use such a targeting approach for all customers in the loyalty program. This includes customers for which the retailer does not have enough observations to run our model. However, what a retailer can do for the customers with too few observations is to try to predict the parameters from our proposed model using customer, category, and store characteristics including the competitive situation around the focal store. We did this using a reduced form of our proposed model (that does not account for endogeneity and neglects the share-of-transactions increase parameter) and predicted the category-level share-of-transactions parameter for customers with too few observations.

We present the predictors and their relative importance in a boosted tree model in Table 7. The boosted tree model⁸ (Friedman 2002) yielded substantially better out-of-sample forecasting accuracy (r²_{out-of-sample} = 0.54) in terms of root mean squared error (i.e., RMSE = 0.089) as compared to the results of an ordinary least squares (OLS) regression (RMSE = 0.168). Our results show that customer-level descriptors, derived from transaction data are of the highest relative importance followed by category-level descriptors. The spatial descriptors regarding the store and its catchment area hardly possessed any predictive power in explaining the share-of-transaction parameter. Table 7⁹ shows that a customer's total purchase value with the retailer, the extent to which the customer seeks variety, her/his shopping frequency at the focal store, and the extent to which a customer can be considered habitual (Liu-Thompkins and Tam 2013) are the most important predictors of the share-of-transaction parameter. Total purchase value and more habitual customers relate to a higher share-of-transactions while the effect of variety seeking relates to lower share-of-transactions (i.e., customers with generally bigger overall interpurchase times and customers that purchase a wide variety of different products tend to have lower share-of-transactions).

Thus, based on the results of this boosted tree model, parameters for customers with too few observations can be predicted. Based on these predictions, the same targeting logic can be applied as demonstrated for the customers with sufficient data for direct model estimation.

Digitalization is one of the key drivers of personalization in retailing. It allows retailers to provide personalized promotions to their customers for those products that they are likely to shop elsewhere just when they enter the store. However, established targeting methods based on observed information such as RFM or recommender systems based on CS fail to incorporate the timing of promotions and neglect customers’ outside potential. In addition, a single retailer observes neither information on a customer’s true purchase timing across retailers nor information on the customer’s purchases at competitors (Gedenk et al. 2010). Knowing about such unobservable purchase patterns would enable the retailer to improve his/her targeting strategy by addressing customers with personalized price promotions in the right product category at the right point in time.

To do so, we suggest a live targeting approach that leverages customers’ purchasing patterns (i.e., interpurchase times) on the category level. Our approach uses observed purchase data of the focal retailer and provides the retailer with information on an individual’s true category-level interpurchase times, share-of-transactions, and reaction to price promotions—accounting for endogeneity arising through purchases on promotion. Retailers can use the category-level insights gained from our model to improve the targeting of personalized price promotions during a store visit.

To validate our model, we ran it on transaction data from a panel provider who tracks purchases on promotion and on transaction data of a grocery retailer who provides in-store personalized price promotions to his/her loyalty cardholders (Appendix 3 and Appendix 4).

We then provided a targeting strategy for retailers to choose the right category-level promotions for their customers when they enter their store: (1) Retailers should use the customer’s true category-level interpurchase time to determine for which categories a promotion would be well timed when the customer enters the store. (2) To select a subset among these well-timed categories, retailers should look at the customer’s share-of-transactions and the expected increase in response to purchases on promotion. At the focal retailer, the results showed that the promotional effect for well-timed categories is highest when the category-level share-of-transactions is lowest. Therefore, among those categories that are well timed when a customer enters the store, we recommend retailers to choose those where the customer’s share-of-transactions is lowest.

For a comparison of live targeting with CS- and RFM-based targeting, we then (ex post) simulated the application of different methods on data from a grocery retailer who actually provides targeted promotions when customers enter the store. Furthermore, we estimated the effect size of RFM-based vs. live targeting-based promotions using propensity score matching. These results show that the reaction of customers to personalized price promotions was significantly stronger when they faced live-targeted promotions as opposed to RFM-based promotions. Matching analysis shows that live targeting has a stronger impact on purchase frequency and revenues than RFM-based targeting (purchase frequency: + 113.74 percent vs. + 69.54 percent; revenue: + 99.28 percent vs. + 56.96 percent). The results illustrate that live targeting can assist retailers in a digital environment in identifying and targeting customers right before a purchase that might have otherwise taken place at the competition.

There are several limitations of our research that might be addressed in future research. First, our model is subject to parametric assumptions. However, the fact that even flexible models with less restrictive assumptions such as the recent nonparametric model of Gabel and Timoshenko (2021) come to similar conclusions, namely the negative impact of a recent purchase in a category on all products within the category, provides support for our parametric assumptions.

The inclusion of more categories, however, increases complexity, and the question arises whether large category-level models should directly incorporate cross-category effects (instead of indirectly via a semi-informed prior) (cf. Jang et al. 2016). Previous research has found that cross-category effects are often small or nonexistent (Gelper et al. 2016; Russell and Petersen 2000). However, recent research (Gabel and Timoshenko 2021; Ruiz et al. 2020) has found promising ways to model the impact of assortment interactions even at the product level. It would be interesting to see future work compare state-of-the-art product-level models like SHOPPER (Ruiz et al. 2020) or the deep learning method of Gabel and Timoshenko (2021) with our proposed live targeting approach.