In February 2014, Golden Orb sponsored the Direct Marketing theatre at the Technology for Marketing and Advertising conference (TFM&A) at Earls Court in London. At the conference, we presented two case studies from our work with Direct Wines. The first of these was entitled ‘Forecasting customer lifetime value – see how Direct Wines addressed this thorny issue’.

This talk set out the principles underlying the notion of customer lifetime value and explained the data requirements and mathematical techniques needed in order to calculate it. In particular, it focused on building models to forecast customer value shortly after acquisition, allowing conclusions to be drawn about the efficacy of different marketing techniques and strategies. The full set of slides can be downloaded here. The key points covered in the talk were:

•  Many direct marketing businesses focus too much on cost per recruit. The advantage of this measure is that it is relatively easy to calculate and available very early on. However, it does not take account of the different quality of recruits acquired through different channels and with different mechanics. What matters most to a business is the difference between average customer lifetime value and cost per recruit. We call this measure added-value per recruit.
• When calculating cost per recruit, it is important to take account of all relevant costs and discounts. Direct Wines use a measure they call ‘Full cost per recruit’ which is the net contribution loss for the campaign divided by the number of recruits. The lifetime value is also calculated at the net contribution level.
• When calculating ‘lifetime value’ it is necessary to determine the time horizon. We use a simple 3-year time horizon, but if a longer view is taken, it is necessary to discount future cash flows.
• We use linear regression to forecast the 3-year lifetime revenue per recruit using a number of predictor variables. The calculations are done by response code and month of recruitment which is sufficiently detailed to provide valuable information, but sufficiently aggregate to be manageable. Forecasting the total 3-year revenue for a cohort of recruits, rather than the revenue per recruit is dominated by the cohort size, masking the effect of good and bad recruits.
• It is generally more accurate to forecast at the highest level possible (e.g. total revenue), rather than forecasting individual revenue streams and adding them together. If a forecast by revenue stream is required, it is generally better to forecast the top level and split it down, rather than vice versa.
• The revenue forecast can be split over future time periods using a simple curve, and broken down to net contribution using simple percentages as trying to be more precise is not warranted by the accuracy of the models.

Using these techniques, it is generally possible to start modelling customer lifetime value after about 4 months and to generate a reasonably good forecast after 6 months. However, the degree of accuracy depends to a large extent on the market and number of recruits, so it is not possible to be precise about how early you can start forecasting. The slides give an indication of how forecast accuracy improves over time.