2008 DMEF Customer Lifetime Value Modeling (Task 2) Zainab Jamal ⁎ & Alex Zhang HP Labs, Palo Alto, CA 94304, USA Abstract We describe our approach for predicting individual donor's total gift amount over a two-year target period. We divide the donors into 8 segments; for each segment, we fit a logit model for predicting the probability of giving, and a log-linear model for predicting the amount of gifts conditional on a donor giving. We found that recency, frequency, and first gift amount are good predictors of the probability of giving, while time- weighted total gift amount in the past years is a good predictor for future gift amount. © 2009 Direct Marketing Educational Foundation, Inc. Published by Elsevier Inc. All rights reserved. Keywords: Lifetime value (LTV); Logit models; Linear models; Recency; Frequency; Monetary (RFM) Our approach involves several sub-tasks: (1) Re-labeling the years to determine a training period and a test period; (2) Segmenting the 21,166 donors and building a separate model for each segment; (3) Looking for predictive attributes and determining the model form; (4) Applying the model to the target period. Below, we describe each of these components in more detail. Training period and test period The observation period is from 2002-01-01 (here we use yyyy-mm-dd format) through 2006-08-31; the “target period” for prediction is from 2006-09-01 through 2008-08-31. Since the target period is 2 years, we use the last 2-year period from 2004-09-01 through 2006-08-31 as our test period; hence, our dependent variable y i is the total gift amount (in dollars) of donor i in the test period 2004-09-01 through 2006-08-31. Shown in Fig. 1, we re-label the years as follows: BY0 = 2002-01-01 through 2002-08-31 (8 months); BY1 = 2002-09-01 through 2003-08-31 (1 year); BY2 = 2003-09-01 through 2004-08-31 (1 year); BY3 = 2004-09-01 through 2005-08-31 (1 year); BY4 = 2004-09-01 through 2005-08-31 (1 year). Observations for donor i prior to 2004-09-01 are used to form potential predictors. In effect, we set 2004-09-01 as our time 0. Recency attribute values will then be computed relative to time 0; for example, a recency of 61 days means that the donor's last gift had occurred 61days prior to 2004-09-01 (i.e. on 2004-07-01). Segmenting the 21,166 donors We divide the donors into one-time donors and multi-time donors (who gave at least twice). The one-time donors gave exactly one gift between 2002-01-01 and 2004-09-01. As an initial investigation, we examine the time pattern of giving of individual donors, see below for some (more prominent) cases, where each chart shows the gift amount (the vertical axis) and the date of each gift (the horizontal axis) of an individual donor. We note that 54.2% of the 21,166 donors gave only once — the pattern typified in the first of the 12 charts below (Fig. 2). We then use a CART tree to segment the 21,166 donors. Combined with our own intuition and judgment, we eventually divide the 21,166 donors into 8 segments, depending on each donor's first gift amount, recency (last gift date prior to 2004- 09-01, here “the last two years” refers to 2002-09-1 through 2004-08-31, and “the past year” refers to 2003-09-01 through 2004-08-31), and frequency (number of gifts): - One-time donors: 11,474 donors. All one-time donors are lapsed donors. Depending on the amount Available online at www.sciencedirect.com Journal of Interactive Marketing 23 (2009) 279 – 283 www.elsevier.com/locate/intmar ⁎ Corresponding author. E-mail addresses: zainab.jamal@hp.com (Z. Jamal), alex.zhang@hp.com (A. Zhang). 1094-9968/$ - see front matter © 2009 Direct Marketing Educational Foundation, Inc. Published by Elsevier Inc. All rights reserved. doi:10.1016/j.intmar.2009.04.004