Merchant attrition has been a problem that the payments industry has been grappling for quite some time now. The industry participants have been exploring the use of analytics to get a better handle on this problem. However, the use of technology has delivered underwhelming results - no significant headway seems to have been made in finding an effective solution till date. Does this mean that the problem is one that cannot be solved by technology, or is the time right for some new thinking in the use of analytics in the merchant attrition context?
A friend of mine owns and runs two popular gyms in my neighborhood. On a recent visit around Thanksgiving, I found him at the counter, looking agitated. Turns out his payment processor had recently switched him over to a newer model of point of sale (POS) equipment the previous month, and his staff was having a hard time trying to get some types of authorizations through. All this led to his customers queuing up at the desk sometimes - leading to a bad overall experience for everyone involved. I asked him whether he had tried speaking to the company about this. He said he had tried various channels, but the company did not seem to be able to solve his problem. I ran into him again a couple of days back, and he told me that he had finally stopped trying and had switched over to a new payment provider earlier this year. This small tale got me thinking - his provider had a three month window between his talking to them about having a problem and his leaving. Surely they would have known that he was thinking of leaving? Surely they could have done something to retain him?
Though a sample of one from the episode above cannot be termed representative of anything, the payment services industry does indeed have a merchant attrition problem. Per some estimates[i], more than 20% of merchants are likely to move from their current provider - ISO or acquirer. 16-17% of this attrition is voluntary, and hence, possibly preventable. Industry consensus also seems to indicate that it takes roughly 3 new merchant clients to make up for each merchant that leaves the portfolio[ii]. The industry spends >$1 billion every year in merchant acquisition costs[iii]. In my friend's case above, that would mean 6 new gyms (or their equivalents) to make up for his lost business - not an easy task in a very competitive industry! These numbers are scary - one-fifth of the portfolio churning every year would give any executive sleepless nights. Not surprisingly, preventing merchant attrition is one of the top to-do items for CXOs in the industry today.
Most conversations on attrition seem to focus on a single aspect - price. It should be noted that price does seem to be the largest reason for merchants to leave a provider. Price reduction was also considered by the processors to be the topmost tool in their retention armory[iv]. It is also equally important to note that price is not the only reason merchants leave. To go back to the example I started with, my friend had told me that he did not have any issue with the contract that he had with his providers - he thought that he was paying a fair price for the services that he was getting from them. His issues lay elsewhere - in fact, he said he was happy with them till he stopped being happy. Also, another fact that the industry agrees with is that their current approach to merchant attrition is mostly reactive. The providers mostly come to know of a merchant leaving when s/he has sent them a request to terminate services. By then, they have already signed the contract with the next provider and are not likely to stay back.
Firms have started utilizing the power of technology to understand the reasons behind merchant attrition, and also try and predict when a merchant is likely to leave. Most of these efforts center on post-facto analysis of why the merchant left, via MIS reports. Some organizations have started using predictive analytics for this purpose too, but I find that these efforts are mostly very restricted - typically, constrained by the data that is used as input to conduct the analysis. The card networks have started offering acquiring banks some of these capabilities too, but they do not have a complete view of the merchant relationship like the acquirers or their processors do. As can be expected, most predictive analytics today is based on segmenting the merchant, and then analyzing their transaction feeds. Any reduction in transaction patterns - volumes or amounts - are thought to be an indicator of the likelihood of their leaving. Though this indicator is an important component in attrition analysis, it is definitely not the only one or even the best one - for example, this indicator might not provide the business sufficient lead time in predicting attrition. The payment processors and their partners today might not be making very good use of the entire gamut of what is possible through analytics - to understand, track and solve this very important problem.
For effective predictive analytics, the possible factors that contribute to merchant attrition need to be first identified. An example of such a factor is price. However, price does not necessarily mean only a dollar figure - it can have multiple nuances. Is pricing high relative to the other merchants in the segment? Is pricing high as a proportion to the merchant's revenues? There are more than 12 such factors can be identified for preliminary analysis. Next, we also need to identify indicators that can be used to analyze whether there is a danger of a particular merchant leaving. The indicators should be tangible and measurable. Merchant transaction patterns are an example of such indicators. There are about 15 such indicators that can be used in this analysis - across merchant transactions and settlements, servicing and other contexts. The factors and indicators form the basis of building hypotheses around why attrition occurs and how it can be detected before the merchant's departure is fait accompli. The hypotheses are created in the form of 'models' - for example, scorecards.
The analytics itself is most likely (in fact, recommended) to be implemented in a phased manner. My experience has been that data availability is always the primary driver behind the decision to follow a phased analytics implementation strategy. Let me explain.
Access to certain 'raw' data required for the analytics is relatively easy - think transaction and settlements data, merchant call center metadata and merchant profile, relationship & pricing data. Such data is structured, relatively easy to access because it is likely being used in reporting already and can be fairly easily 'piped' to the analytics infrastructure. We are looking at around 25 attributes participating in analytics in the initial phases, which are derived from around 80 easily available 'raw' data attributes.
The data that is not likely to be easily available currently is typically unstructured and / or available from external sources. Think of online chat text, contact notes in Salesforce, merchant call center transcripts and sources like Twitter, Facebook or popular online SME forums. Ad hoc surveys being done by providers would also fall in this category. Analytics based on this data can be 'bolted on' in subsequent phases, when the reality of data availability has had an opportunity to catch up with the analytics strategy.
A phased approach to attrition analytics helps the organization start getting quicker benefits from data that is available today, without having to wait for 'data utopia' before these benefits kick in. In the example that I started with, the fact that my friends calls to their service center had increased in frequency would have been detected in the earlier phase if such an approach were to be followed. The fact that he was not particularly pleased would have been detected in the later phase, let's say, using agent notes or other methods. However, the detected change in his calling pattern would probably have been enough to warrant an investigation, along with the other parameters in his score like a change in his transaction patterns.
Once we have the data in place, it's time to decide the analytics approaches that will be used on this data to determine likelihood of merchant attrition. Of course, regression analysis to determine co-relation of the indicators with the likelihood of the merchant leaving is almost a given. Most analytics toolsets in the market offer the capability. This capability can be used to build the model / scorecard
Interesting approaches like pattern anomaly detection - of transaction, settlement, engagement and service request patterns - can be used on the data available in the initial phase, to feed into this model. To reiterate, this data is typically easily available and in structured format - and hence, quickly usable.
For example, if the transaction frequency from a merchant reduces drastically along with a corresponding increase in her calls to the service center and then a drop-off in call volume too, is it likely that the case should be investigated for probable attrition?
Now, it might seem intuitive that since the information is already available with the call-center, it might be also available to the retention / relationship team. But having seen the typical functional (and data) siloes in any organization, I can assure you that this is not a given. This way, the analytics route also becomes a (arguably, not very efficient) mechanism to improve data flow across teams, till the process kinks get worked out.
The next phase of analytics can be even more interesting. Here we are looking at using approaches like entity resolution, network analytics, web and voice / sentiment analytics to leverage data that is unstructured and often 'outside the enterprise' in our analytics equation.
Entity resolution along with network analytics can help the provider identify each specific merchant across multiple channels (including 3rd party ones like Twitter and web forums) and also gauge how much they might influence attrition other than their own (and also acquisition of other merchants).
Voice, text and sentiment analytics can look at data from voice recordings, call transcripts and contact notes to understand the 'tone' of the merchant's communication with the organization, and then further use that tone as one of the inputs in predicting the likelihood of attrition.
These slightly more (arguably) exotic approaches can be implemented using point solutions which are built for these specific purposes. However, in my experience, understanding the implications of the output from such solutions and then leveraging them effectively in the overall attrition analytics model might require longer lead times as compared to the more 'traditional' analytics approaches.
However, their use can definitely lead to a stronger merchant retention program, as compared to one that uses just the traditional insight generation mechanisms.
The next step from that point onwards is a move towards the holy grail of any analytics program today - prescriptive analytics. A program which not only tells the organization which merchants are likely to leave, but also goes ahead and suggests exactly what needs to be done by the relationship, service, product and other teams to retain them. But that's grist for another post.
Right now, the analytics crystal ball seems to show a future full of happy merchants!
- FS Solutions Group