Artificial intelligence spots who is likely to buy
A new prediction tool from @RISK helps identify customers ripe for up- and cross-selling. It is based on the same technology as @RISK’s existing Pathfinder, which identifies those ready to defect.
@RISK, Inc., developer of artificial intelligence systems and data mining applications that can predict and shape customer behaviour, has introduced “not @RISK”, a prediction tool that will help identify those customers that are ripe for cross-selling and up-selling. The technology is a direct by-product of the @RISK customer behaviour prediction systems that have been successfully used to identify those customers that are most likely to defect.
Neural networks @RISK uses neural network protocols and causal inference algorithms to search through customers’ transactional data. These detect patterns and trends in their transactions and, when their patterns begin to diverge from what is “normal” for that individual and begin to mirror those of cohorts who have defected in the past, the system flags the individual as at risk. The modelling protocols take into account the fact that this pattern may well vary by segment and may in fact change, based on market conditions.
Correlate or indicate? Systems that predict defection have to be able to distinguish between variables that simply correlate with defection and those that actually indicate defection. By using artificial intelligence and machine learning techniques, @RISK’s programmes can filter out spurious correlators and focus on indicators. This leads to better long term accuracy and precision than methods like logistic regression, CHAID or CART can provide.
Not Mr Average Recognising that forecasting what the average customer will do is of little help, @RISK uses a Jordan (or Hierarchical Mixture of Experts) Network. This neural network architecture simultaneously finds the latent classes underlying the transactional data and fits separate non-linear models to each class. This latent class, non-linear regression (provided by the Jordan Network), is based on a set of predictors with substantive rather than spurious correlation and leads to higher accuracy and stability over time.