5.1 Data description
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.
Table 3
Descriptive statistics including mean and standard deviation of the retailer data set (estimation period: 05/2015 until the end of 06/2016)
Banana | 1268 | 19.0 (14.6) | 0.7 (0.3) | 1.2 (0.5) | 10.8 (9.0) | 0.41 (0.49) | 21.4 (13.8) |
Tomato | 952 | 17.5 (11.3) | 1.0 (0.3) | 2.3 (0.7) | 10.8 (8.1) | 0.21 (0.41) | 22.6 (13.4) |
Bread | 758 | 23.2 (17.4) | 1.2 (0.3) | 1.6 (0.5) | 14.4 (11.4) | 0.23 (0.42) | 19.3 (12.3) |
Butter | 561 | 19.8 (12.3) | 1.4 (0.4) | 1.9 (0.7) | 11.4 (9.0) | 0.25 (0.43) | 20.5 (11.6) |
Coffee | 85 | 17.4 (10.9) | 1.8 (0.8) | 5.1 (2.9) | 7.2 (5.2) | 0.39 (0.49) | 24.5 (15.0) |
Pizza | 332 | 14.1 (9.9) | 1.6 (0.6) | 3.8 (1.6) | 21.0 (16.0) | 0.18 (0.38) | 25.8 (12.9) |
Milk | 398 | 27.4 (18.5) | 1.6 (0.8) | 1.5 (0.9) | 5.8 (4.4) | 0.38 (0.48) | 16.4 (10.1) |
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.
5.3 Comparing our proposed targeting approach to an RFM-based targeting 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.
Table 4
Summary of customer selection strategies
All promotions at the focal retailer were generated using cosine similarity | Target the top 98 or 490 customers with the highest RFM score in a category at the end of the selection period | From all the customers, who have received a well-timed promotion in a category, target the top 10% (i.e., 98) or the top 50% (i.e., 490) who have the lowest share-of-transactions during the selection period (highest outside potential) |
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.
Table 5
Comparison of different targeting approaches: CS, RFM, and live targeting
Redemption rate of promotions assigned under the approaches (A) | 0.277 | 0.255 | 0.424** | 0.297 | 0.338* |
Promotional characteristics |
Selected customers per category (S) | 490 | 98 | 98 | 490 | 490 |
Promotions assigned under the respective approach (A) | 1727 | 353 | 118 | 1,914 | 693 |
Sample characteristics |
RFM score (S) | 8.822 | 14.031 | 7.653 | 12.710 | 8.163 |
Share-of-transactions parameter (without increase, pi) (S) | 0.491 | 0.610 | 0.219 | 0.607 | 0.250 |
Share-of-transactions increase parameter (without decay, δi) (S) | 0.211 | 0.149 | 0.314 | 0.169 | 0.392 |
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.
5.4 Revenue and purchase frequency analysis using propensity score matching
As a second illustration of the power of live targeting vs. RFM-based
7 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.
Table 6
Results of the revenue and purchase frequency analysis using propensity score matching
Purchase frequency (A) | 2.304 | 1.359 | + 69.54% (+ 0.945) | 2.909 | 1.361 | + 113.74% (+ 1.548***) |
Revenues in € (A) | 3.505 | 2.233 | + 56.96% (+ 1.272) | 4.412 | 2.214 | + 99.28% (+ 2.198***) |
Sample characteristics |
N | 973 | 973 | | 817 | 817 | |
RFM score (between 3 and 15) (S) | 9.199 | 8.358 | | 9.261 | 8.285 | |
Recency score (between 1 and 5) (S) | 2.785 | 3.022 | | 2.542 | 3.056 | |
Frequency score (between 1 and 5) (S) | 3.195 | 2.602 | | 3.365 | 2.540 | |
Monetary value score (between 1 and 5) (S) | 3.219 | 2.734 | | 3.354 | 2.689 | |
Share-of-transactions parameter (without increase, pi) (S) | 0.509 | 0.434 | | 0.549 | 0.437 | |
Share-of-transactions increase parameter (without decay, δi) (S) | 0.202 | 0.226 | | 0.193 | 0.227 | |
Rate parameter of the Erlang distribution (βi) (S) | 0.281 | 0.320 | | 0.283 | 0.309 | |
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.
5.5 Guidelines when implementing live targeting
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 model
8 (Friedman
2002) yielded substantially better out-of-sample forecasting accuracy (
r2out-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).
Table 7
Relative importance of predictors used in boosted tree model
Total Euro value of a customer’s purchases with the retailer | Individual | 16.4 |
Mean variety seeking of a customer | Individual | 14.5 |
Overall purchase frequency of a customer with the retailer | Individual | 14.0 |
Degree of habituality of a customer | Individual | 10.8 |
Time enrolled in loyalty program after its launch | Individual | 4.3 |
Mean price point of a customer | Individual | 4.1 |
Share of baskets with prior system usage | Individual | 4.0 |
Number of promotion kiosk uses | Individual | 4.0 |
Average discount offered to the customer | Individual | 3.7 |
Total savings of a customer based on kiosk system | Individual | 3.1 |
Average redemption rate of a customer | Individual | 2.6 |
Average basket size of a customer | Individual | 2.6 |
Number of different categories a customer purchases in | Individual | 2.4 |
Number of baskets of a customer | Individual | 2.0 |
Coefficient of variation in price within a category | Category | 1.2 |
Number of free product rewards redeemed | Individual | 0.9 |
Average quantity purchased per category | Category | 0.7 |
Number of stores of the retailer a customer purchases in | Individual | 0.6 |
Number of competitors close to the focal store of a customer | Spatial | 0.5 |
Coefficient of variation in category interpurchase time | Category | 0.5 |
Population heterogeneity within district | Spatial | 0.5 |
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.