Predictive analysis in Armstrong One
The 24 different customer journeys of Armstrong One are based on a wide array of analytical models. Each journey is designed to fit a specific purpose and all elements of the modelling have been optimized accordingly. In addition to providing predictive models, the analytical processes of Armstrong One automatically consider rules for timing, triggering, cropping and removal of data.
Analytical Models with purpose specific complexity
The complexity of the models ranges from traditional scoring models to combinations of churn- and association models. In accordance to their purpose, some journeys are limited to entail basic analytical methods.
For most customer journeys, behavioural data is applied to perform association analyses of sales’ data for different consumer profiles and the majority of the models are automatically improved over time by input from a closed data feedback loop in Armstrong One and from all other available sources. That means that the predictions improve as the model is exposed to the customer base and that the model will not expire. In addition, the Armstrong One team constantly develops and improves the models.
The largest family of predictive models are used to perform association analysis. The most common algorithms for this are Apriori, Carma and Sequence. These models are valuable across a variety of industries, but particularly useful within retail.
Traditionally, simple logical rules as causality – “customers who purchase X also buy Y” are often used even within digital marketing. However, the advanced algorithms within Armstrong One automatically identifies these co-relations automatically, efficiently and with far more flexibility by e.g. considering the lift given to sales from one product to another. Including these methods, allows Armstrong One to identify co-occurring items, cross-sales, substitutes and complementary products.
Classification models, e.g. logistic regression and classification trees, are yet another important family of predictive models. These may predict the occurrence of an event as well as being able to calculate the probability of that event taking place. Classification models have been implemented together with a time series model in the churn models in Armstrong One. These are then able to predict if a customer is likely to churn after a certain time period, based on his or her behaviour. The same technique can be used to identify which type of customer is likely to participate in a sweepstake.
All models are trained on suitable segments so that significant associations can be identified. However, the models are always applied on an individual level. Consequently, the resulting targeting rules are a result of a suitable mixture of the consumption habits of the individual customer and the buying behaviour of other customers.