The ability to change or otherwise influence the perception of your products is foundational to retention marketing. However, while all consumers who demonstrate a decline in engagement may be considered targets for retention marketing, identifying persuadable consumers as targets will yield dramatically greater returns.

At Newsday, we built a comprehensive customer engagement monitoring plan consisting of more than 10 legacy and digital customer touchpoints with our brand. It was aimed at identifying each individual customer’s habits of engagement and deviations from such. In analytical terms, this process is defined as churn modeling.

While our churn modelling identifies at-risk customers (those with low or declining engagement), it’s our persuasion modelling (also known as “net-lift modelling”) that identifies both the customers who are open to being persuaded to stay and the optimal marketing treatment to persuade them.

As part of our efforts to enhance our relationship with subscribers and persuade them to stay, we initiated a surprise-and-delight programme that provides for giving various gifts and gratitude to subscribers with different levels of likelihood to churn.

Among the different treatments, we found high-churn subs (those with high probability to churn, according to our churn model) responded best to a significant gift such as a portable smart device charger. Their churn rate was reduced by 40% at five months post-marketing.

Since we were able to reduce churn significantly among this group with a gift, high-churn subs could be a good target. However, a significant gift means significant cost, so we want to take this one step further to identify subgroups of targets who would respond better to different and less costly treatments. To achieve this, we built a net-lift model to forecast the true incremental impact of treatment for each different subscriber.

Net-lift models are based on classifying subscribers into four groups:

1. Sure things: Already persuaded and will be likely to stay without intervention.
2. Lost causes: Already ready to churn and not likely to be persuaded.
3. Self-selectors/don’t disturb: Like to make decision on their own and retention efforts could have an adverse effect.
4. The persuadable: Will have a positive response to retention marketing. Need some convincing and open to being convinced. These are the target customers.

The only segment that provides true incremental response is the persuadable, and it should always be our target. To classify the persuadable, a net-lift model is built using a combination of variables, including demographic, geographic, and engagement history from both internal and external sources. This helps to identify a subgroup that will respond if given a particular treatment and will not respond if not treated.

In this application, net lift is the amount of churn rate reduction observed in a test group versus a control group. Therefore, the modelling sample is made up of some treated individuals and some untreated ones with similar characteristics. The model proceeds by subsetting subscribers based on different features to maximise the difference between the test group and control group (i.e. net lift).

Here, a tree-based algorithm is implemented as illustrated below.

For all subscribers, the net lift provided by retention efforts is only 1%. It’s important to note here that this is an averaged effect of some subs responding positively and others responding negatively or not at all.

After splitting subscribers based on different features, we can find a better target — subscribers with both characteristic A and characteristic C — and their net lift score can be enhanced to 5%. Splitting would be continued, until net lift cannot be enhanced or the size of the subgroup is too small to be useful.

From here, we implemented a random forest algorithm as a way of averaging multiple deep random decision trees, thus greatly reducing the variance and boosting performance.

Sample variability is introduced into the random forest by sample bagging and features bagging (i.e. using a different subset of respondents and a different set of features, respectively, for each tree), and the final net lift score is outputted from the average prediction of all random decision trees.

Utilising our net-lift model, we successfully identified a group of persuadable high-churn subscribers that would respond well even to the most economical treatment — greeting cards. We implemented the strategy, and the churn rate of identified persuadable subscribers was reduced by 58% six months post-marketing of the greeting card.

Another application of net-lift modelling at Newsday was to identify the best target for our Thursday upgrade programme — adding a Thursday newspaper to Sunday-only subscriptions on a no-charge, opt-out basis.

When we executed this programme in past years, our success measurement was focused on circulation volume gains and absent of impact on retention.

Contrary to our thoughts regarding impact on retention, in aggregate, the no-charge Thursday opt-out actually created greater churn. However, it was likely to generate greater retention for some subscribers. Therefore, we built a net-lift model to identify the persuadable subs who would respond positively to a Thursday upgrade.

Based on our model, about 40% of Sunday-only subscribers with five key characteristics are identified as the persuadable, and they will be upgraded later this year.

In conclusion, while it is important to understand each individual consumer’s engagement habits with the various touchpoints of your brand via churn modelling, it is persuasion modelling that identifies both the targets among those with high or declining engagement and the respective optimised treatments.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .