In this section, we will present the CRM implementation at IBM (a B2B company), from initial model development to actual field study implementation. The whole idea of the case study is to showcase how CLV was used as a useful indicator of customer profitability for IBM's marketing decisions toward maximizing the firm's value [2]. This study shows that CLV performs better in measuring customer value as compared to IBM's customer selection metric.
By implementing CLV measurement and maximization on about 35 000 customers, IBM was able to increase its revenue 10-fold without any changes in the level of marketing investment. This increase in revenue was made possible just by reallocating the marketing communication resources to the right customer groups.
IBM, a multinational technology and consulting service provider to B2B customers, aimed to maximize its overall profitability through customer-centric CRM strategies. Specifically, IBM sought to maximize its profitability by prioritizing its resource allocation to the most profitable customers. To measure customer profitability, it used Customer Spending Score (CSS), which was defined as the total revenue that can be expected from a customer in the next year. Following the scoring, IBM then ranked customers into 10 deciles according to their corresponding CSS values. The top one or two deciles were then targeted and allocated more marketing resources.
In 2004, IBM felt the need to move to implementing a better indicator for customer value measurement than CSS, because CSS focused primarily on customer revenues and largely ignored the variable cost of servicing customers. CLV was then proposed as an alternative indicator, and a CLV-based management framework was suggested for implementation for IBM's customers to improve their profitability.
Specifically, IBM wanted to test the following belief: ‘When all other drivers remain unchanged, can an increase in the level of contacts with the right customers create high value from the low-value customers?’ To accomplish this objective, a CLV-based management framework was adopted to design customer management initiatives, as illustrated in Figure 8.2.
A two-stage process was used to develop and implement the CLV management framework at IBM. In the first stage, several models were developed to generate inputs for implementation. In the second stage, a field study was conducted based on recommendations from the models developed in the first stage. The aim of the field study was to test IBM's main objective – whether an increase in contacts to the right customer creates high value from low-value customers when all other drivers are similar. Figure 8.3 illustrates the two-stage approach involved in implementing a CLV management framework at IBM.
Data on 20 000 mid-market companies of IBM were used in this study and the subsequent implementation. Mid-market companies are defined as companies that have the number of employees within the range of 100 and 999 at the enterprise level, and with total enterprise revenues to IBM from 2001 to 2003 that were over $25 000. The data were collected at the monthly transaction level, and per establishment. Each enterprise can have more than one establishment (e.g., different locations making independent decisions). In total, 2.5 million observations were collected (72 months each for the 35 131 establishments).
The model development stage involves developing a framework to measure individual customer's CLV and selecting the right customers, propose an optimized resource allocation and contact strategy, and build propensity models. Specifically, this stage involves the following three phases:
In this phase, the always-a-share approach for measuring CLV was adopted because of its relevance to the non-contractual setting of IBM. This approach assumes that customers never ‘quit’ their relationship with a firm; rather they demonstrate only dormancy in their relationship with the firm while transacting with other firms and always have a probability (no matter how small) of purchasing in a given time period [3]. When they return to the relationship, they retain the memory about their prior relationship with the firm. Accordingly, the CLV was measured using the following formula:
where:
= lifetime value for customer i
= predicted probability that customer i will purchase in time period j
= predicted contribution margin provided by customer i in time period j
= average cost of a single marketing contact per customer, assumed to be $7 in the study
= number of contacts directed to customer i in time period j
j = index for time periods, months in this case
T = the end of the calibration or observation time frame
r = monthly discount factor, 0.0125 in this case (amounting to a 15% annual rate)
In Equation 8.2, the computation of CLV involved predictions of three aspects: (a) the level of marketing contacts directed toward customer i in time period j (MT), (b) the probability that a customer would purchase in each time period (p(Buy)), and (c) the contribution (in $) provided by the customer in each time period (CM). A description of the prediction of these three aspects is provided first here.
The marketing contacts allocated by the firm toward customer i are determined by the following model:
where x1ij, β1, α1i, u1ij are, respectively, a vector of predictor variables, a vector of corresponding coefficients, an individual-level intercept, and an error term.
It was assumed that customer i purchases from the firm only when the customer's latent utility for purchasing from the firm () exceeds a certain threshold, set to zero in this case. In this study, only the binary outcome variable regarding whether or not the customer purchased in time period j was observed. Consequently, the latent utility of the customer was not observed. The latent utility is mapped to the binary outcome variable (Buyij) as follows:
(8.4)
The latent utility, (), for customer i to purchase from the firm in time period j was then modeled as a function of predictor variables in a linear model:
where, similar to Equation 8.3, x2ij, β2, α2i, u2ij are, respectively, a vector of predictor variables, a vector of corresponding coefficients, an individual-level intercept, and an error term.
We assume that a latent variable, , represents the amount spent by customer i in time period j, irrespective of whether it is with the firm, as a function of predictor variables with a linear structure
where, similar to Equations 8.3 and 8.5, x3ij, β3, α3i, u3ij are, respectively, a vector of predictor variables, a vector of corresponding coefficients, an individual-level intercept and an error term.
If the customer purchased from the firm in time period j, then the firm observed the contribution margin provided by the customer as follows:
(8.7)
The three aspects involved in the computation of CLV are inherently correlated. The level of marketing contacts directed toward a customer depends on customer characteristics, past customer behavior, and the past level of marketing resources allocated to the customer. The probability that a customer would purchase is likely to be dependent on the level of marketing resources directed toward the customer, and, finally, the customer provides profits to the firm only if a purchase is made. In the modeling framework used in this study, these aspects of firm and customer behavior were allowed to be correlated with each other. To jointly model the marketing contacts, the probability of purchase, and the contribution margin, a model structure based on a ‘seemingly unrelated regression’ (SUR) was used. The likelihood that summarizes our model structure is provided below:
where is customer i's latent utility for purchasing in time period j.
The covariance structure of the errors in Equations 8.3, 8.5, and 8.6 was modeled as
(8.9)
Such a method of modeling provided the possibility of correlations among the three residuals. Further, σ11 was fixed to be equal to 1 to ensure model identification. The covariance structure of the errors accounted for any unobserved dependence between a firm's decision to contact a customer (MT), a customer's decision to purchase from the firm (Buy), and the amount of money the customer spends with the firm (CM). By allowing β = [β1, β2, β3] and α = [α1, α2, α3], the simultaneous equation model gives rise to the likelihood specified in Equation 8.8 – the model likelihood equation. The customer-specific intercept terms were obtained from a multivariate normal distribution as follows:
where:
Zi = a p × 1 vector of customer characteristics
Δ = a 3 × p matrix of coefficients for the customer characteristics
= a 3 × 3 variance–covariance matrix
p = number of customer characteristics that are used to capture heterogeneity
Using the modeling approach described above, the first 54 months of data were used to estimate model parameters, and the CLV score was computed for each customer for 36 months (2002 through 2004).
In calculating the CLV score, the drivers of the respective parameters, such as marketing contacts, probability of purchase, and contribution margin at IBM, were identified. The interactions between the drivers were also evaluated so that the CLV modeling was conducted along four critical dimensions: (1) modeling firm decisions, (2) providing a forward-looking cost allocation strategy, (3) imputing missing contribution margin, and (4) accommodating unobserved dependence among levels of marketing contacts, purchase incidence, and contribution margin.
The drivers of marketing contacts were past customer spending, past levels of marketing contacts, cross-buying and recency, and past purchase activity. The drivers of probability of purchase and contribution margin were categorized into customer relationship characteristics and customer firmographics. The customer relationship characteristics included drivers such as past customer spending level, cross-buying behavior, purchase frequency, recency of purchase, and past purchase activity, and the marketing contacts by the firm. The customer firmographics included drivers such as sales of an establishment (a measure of the size of the establishment), and the installed level of PCs in the establishment (a measure of the level of demand for IT products in the establishment).
The coefficient estimates of the drivers of CLV are reported in Table 8.2. The values are the posterior means and variances. A parameter is considered not significant if a zero exists within the 2.5th percentile and 97.5th percentile values of the posterior distribution for that parameter.
Estimation results | ||
Dependent variables | ||
CLV | The discounted value of all expected future profits, or customer lifetime value | |
Coefficients | ||
Independent variables | Mean* | Variance* |
Level of marketing contacts | ||
Lagged level of contacts | 0.7366 | 0.0316 |
Two-period lagged level of contacts | 0.3239 | 0.0333 |
Lagged average number of purchases | 0.5836 | 0.2158 |
Two-period indicator of purchase | 4.7114 | 1.4023 |
Interaction of cross-buying and recency | −0.0149 | 0.0035 |
Lagged contribution margin | 0.6016 | 0.1128 |
Purchase incidence | ||
Lagged indicator of purchase | 0.6573 | 0.0902 |
Two-period lagged indicator of purchase | 0.2172 | 0.0891 |
Lagged average level of contribution margin | 0.0056 | 0.0026 |
Log of lagged level of contacts | 0.0041 | 0.0012 |
Interaction of cross-buying and recency | −0.0047 | 0.0074 |
Interaction of log of lagged level of contacts and lagged indicator of purchase | 0.0004 | 0.0002 |
Contribution margin | ||
Lagged contribution margin | 0.8612 | 0.0247 |
Lagged average contribution margin | 0.7442 | 0.0325 |
Cross-buying | 0.2858 | 0.1075 |
Frequency of purchases | 7.3692 | 1.887 |
Log of lagged level of contacts | 0.079 | 0.0105 |
Interaction of cross-buying and recency | −0.038 | 0.0565 |
* Mean and variance are computed using the 5th through 95th percentiles of the posterior sample. |
According to Table 8.2, the following key observations were made:
The CLV score for each customer was computed using information from the predictions of marketing contacts, purchase incidence, and contribution margin, as well as the unit marketing costs for each channel. In all, 72 months of historical data were available for model development. Traditional metrics were also computed using the first 54 months of data. Customers were then rank ordered based on the CLV measure as well as on the traditionally used metrics. The comparative performance of the customers (i.e., the observed profits provided by the customers in the last 18 months) in the top 15% of each list of metrics clearly shows the power of CLV to identify the best customers for future targeting (see Table 8.3). Contrary to prior findings, this study found that in non-contractual settings, at least with regard to selecting high-potential customers for future targeting, current profit performs worse than estimates of future profitability.
Using the CLV measurement and the subsequent customer selection approach, a forward-looking cost allocation strategy was implemented on IBM's customers. The cost allocation strategy involved used an optimization algorithm to maximize the sum of expected CLVs for all the customers, as the frequency of marketing contacts for each customer were allowed to be varied. This method facilitated identification of the optimal level of marketing contact for each customer that would maximize the sum of expected CLVs of all customers.
The optimization algorithm was conducted on 5000 customers, with the following parameters: population size = 200, probability of crossover = 0.8, probability of mutation = 0.25, and convergence criteria = difference in optimal solution over the last 10 000 iterations is less than 0.1%. The output from this optimal resource allocation model allowed for the determination of the optimal number of contacts for each customer and the corresponding marketing resources to be allocated to the customer.
The output from the optimal resource allocation model produced the input to the decision-making process about the number of contacts in each channel for each customer in various customer segments. The differences in suggested optimal contact frequencies across various customer segments are shown in Figure 8.4.
This figure shows the contact strategies recommended by IBM's CSS indicator and recommended by the CLV management framework respectively. It is important to note that the CLV framework suggests a rather different strategy. For example, customer segment D, which was suggested to receive the lowest level of contact by CSS indicator, should actually receive the highest level of contact by CLV model.
Based on the CLV contact approach, an optimal contact strategy for the customer segments involved in this study was developed. This strategy involved classifying the contact strategies into four buckets along the CSS and CLV metrics. Table 8.4 illustrates the optimization strategies.
CLV | CSS | |
Low | High | |
High | Direct mail/telesales/catalog/e-mail: | Direct mail/telesales/catalog/e-mail: |
Current interval: 4.82 d | Current interval: 6.3 d | |
Optimal interval: 1.9 d | Optimal interval: 5.3 d | |
Gross value: | Gross value: | |
Current value: $10 936 | Current value: $53 488 | |
Optimal value: $17 809 | Optimal value: $90 522 | |
Low | Direct mail/telesales/catalog/e-mail: | Direct mail/telesales/catalog/e-mail: |
Current interval: 9.7 d | Current interval: 8.4 d | |
Optimal interval: 12.6 d | Optimal interval: 8.3 d | |
Gross value: | Gross value: | |
Current value: $743 | Current value: $1091 | |
Optimal value: $1203 | Optimal value: $2835 |
As shown in Table 8.4, it was recommended that the contact interval through direct mail/telesales/catalog/e-mail to the Low CSS–High CLV and High CSS–High CLV customer segments be increased from 4.82 to 1.9 days and from 6.3 to 5.3 days, respectively. In other words, these customers will have to be contacted more often than the current practice. This increase in contact level would provide an increase in gross value of around 63 and 69% respectively. The biggest lift in gross value is observed in the High CSS–Low CLV group segment, approximately 160%, if IBM were to reduce the current contact interval from 8.4 to 8.3 days. Finally, the Low CSS–Low CLV segment customers should be contacted less frequently, thus less marketing resources should be allocated to this group, which will result in a 62% increase in gross value.
Following this, to optimally reallocate the marketing resources, and to test/assess whether the different resource allocation strategy suggested by the CLV model will translate into higher profits for the firm, a field experiment was conducted. Before conducting the field experiment, the study sought to find out what products to pitch to which customers.
In this phase, the predicted purchase probability for a category, a product for each customer was identified. This allowed IBM to determine which products to pitch to the targeted customers. The rationale behind this phase was the assumption that the customer's need for a certain product type and familiarity with the focal categories are the key drivers of the customer's product category choice. The drivers of product category choice that have significant influence include: (1) the proportion of the same category purchases, that is, the dominance of a category over others; (2) the depth of same-category purchases measured as the number of products purchased within the focal category, that is, knowledge of the focal category; and (3) the breadth of same-category purchases measured as the number of different product types purchased within the focal category.
To determine a customer's propensity to purchase each of the product categories, propensity models were built to predict the propensity of buying products within the following categories: hardware, software, and services. A product message was used in a marketing contact to a customer if the predicted purchase propensity for the category is greater than 0.5 for that customer. The results of the propensity model, together with the CLV model, allowed IBM to convey the right product/service message to the right customers.
The model implementation stage involved implementing the customer selection and resource allocation strategies proposed in Stage 1 in a field study on selected IBM customers. The specific objectives of the field study were to:
To determine the level of resources to be allocated to the respective groups of customers, 35 131 customers were selected and divided into two groups, the Contacted by 2004 (customers who have been contacted previously by IBM until 2004) and the Not Contacted until 2004 (customers who have not been contacted until 2004). In each group, the customers were then rank ordered into deciles, according to their respective CLV scores, as shown in Table 8.5.
As can be seen from this table, customers in decile 10 in the Contact group were not profitable. It was recommended that resources from this group be reallocated to customers in the Not Contacted until 2004 group in deciles 1, 2, 3. Customers in deciles 1–3 of the Not Contacted until 2004 group were identified as those having high purchase propensity for at least one of the three product categories. Such a reallocation implied that customers with higher CLV will be given higher priority to be allocated resources first. The level of resources to be allocated was determined based on the optimum contact strategy described in Phase 2 (Table 8.4). This reallocation process resulted in some customers in decile 1, all customers in deciles 2 and 3 in the Not Contacted group being allocated marketing resources for 2005.
As a result of an improved targeting strategy, the revenue of the Not Contacted until 2004 group increased 10 times in 2005 compared to revenues in 2004. The lift in revenues for the Not Contacted until 2004 but Contacted in 2005 group of customers was about $19.2 million. The incremental revenue due to resource reallocation (after adjusting for the annual growth in customer revenue) among the Not Contacted until 2004 but Contacted in 2005 group of customers was about $19.1 million.
This incremental value was derived from two sources: (a) $7.6 million (nearly 40%) was obtained from the increase in purchase amount from customers who were active in 2004, and (b) $11.4 million (nearly 60 %) was obtained from the reactivated customers (about 273 customers) who were dormant in 2004. Therefore, the average increase in revenue from reactivating dormant customers was about $41 758, and the average increase in revenue from existing customers was about $4160. The effectiveness of this CLV model was reflected in the superior performance of the sales revenue metric. The improved profitability for IBM was made possible by the successful implementation of the CLV-based strategies.