The case of the misleading model
Qbase increases customer acquisition rates with propensity modelling
Qbase were asked by a company in the B2B catalogue market to improve their customer acquisition rates which has been reducing for the past 18 months. At the time, the company were working with another data bureau and were using a uni-variate profile model to select prospects to mail.
A uni-variate model looks at the significance of a variable [for example industry or company size], in isolation to other variables and then says whether or not it is "significant". This approach fails to take into account the significance of all the measureable variables when combined.
For example a uni-variate model might tell you that retail and distribution industry sectors are important. But what you really want to know is that retail and distribution is important... as long as those companies have between 100 and 250 employees and are based in 4 metropolitan areas. Only multi-variate modelling can tell you this by measuring the significance of every single variable in relation to every other variable, not looking at things in isolation.
Qbase produced a detailed response analysis, incorporating Cost per Acquisition and Lifetime Value of each customer acquired. This enabled us to rank each prospect list that had been used in the past three years. The prospects lists with the best performance were combined with a B2B compiled database to create a prospect universe to select data from.
We then enhanced the customer file with demographic variables including SIC [Standard Industrial Classification], number of employees and financial data and produced a multi-variate model based upon the best customer segments.
Golden Rule of modelling: only model your top customer segments, since you only want to find new customers who look like your current best customers.
This exercise was repeated and a new model was created, based upon the highest ranking prospect lists to compare the results with top customer segment model - if the results were similar, the model would be robust and a good predictor of likely high value customers from the prospect base.
The two models were compared and the results showed similar patterns so both models were combined to produce a single model.
The model was then applied to a compiled database of the UK Business Universe and each prospect was scored and ranked according to their propensity to share the same characteristics as the best customers and converted prospects.
3 years on and Qbase continue to refine the model on a quarterly basis and it is used as the sole means of selecting data for mailing. The model has also been applied to a number of other compiled databases. Selections based on the propensity model have have returned up to 2.5 times the ROI of non-modelled prospects.
Customer acquisition has increased and is above plan for the last two years, as is average new customer value.
Not all models are equal, multi-variate modelling is the only way to identify all the possible permutations of attributes likely to look like your best customers or donors.
To find out more about our modelling services, contact Paresh Patel.