Using Sentiment Analysis For Reputation Risk Management

Data Driven Reputation Risk Management

When it comes to risk management, one of the major pillars to monitor and manage is reputational risk. Unlike credit or interest rate risk, reputational risk is hard to define and even harder to quantify. Given the number of discussions on various news platforms, social media channels, and forums, there are hundreds and usually thousands of discussions taking place without a bank’s knowledge. Sites like Yelp, Glassdoor, Twitter, Facebook and similar usually get the bulk of those mentions, but also there are a handful of blogs that are either likely to talk about your bank from a community service perspective or from an investment perspective.  In this post, we explore one way that we monitor and manage reputational risk and walk through our preliminary framework in order to help other banks evolve their thinking.

 

CenterState Bank As A Case Study

 

The larger and more relevant your bank is, the apter the public is going to talk about you. Banks that have active social media programs do so in part to increase that relevancy and in part to attempt to influence that conversation. However, even if your bank has no social media program, your reputation is still at risk every day.

 

For non-exchange traded banks, most of the discussion is usually around a bank’s retail service level or charity work.  This is the bulk of our conversations and is fairly straight forward. However, if you are publicly traded, then there are likely dozens of posts every per day that your bank not only needs to watch but needs to analyze for trends, compliance and changes to reputation.    

 

Sentiment Analysis

 

Last year, for example, CenterState, and our stock symbol “CSFL,” was mentioned in approximately 9,000 separate posts over the course of 12 months (see above for our daily breakdown). About 70% of these posts were related to our equity. That is an average of about 25 posts per day and breaks down to almost six threads (a thread is a conversation between two or more parties) per day talking about your earnings potential, stock price, management or historical equity performance. These are likely conversations that your bank may not be aware of and if they are, may not have a way to quantify.

 

Even outside of retail and equity conversations, banks still generate a fair amount of newsworthy press because of their sponsorship of events, earning releases, M&A work or new products. As can be seen in our analysis below, the bulk of our conversations at CenterState are generated by news outlets. Here, a simple Google search alert can let you know any time your bank is mentioned, but you still have the challenge of trying to ascertain the news article’s impact on reputation.  Making matters more complicated, as you can see below, a fair number of mentions take place on social media and blogs. 

 

Sentiment Analysis By Source

 

The Challenge of Quantifying Reputational Risk 

 

Any news story about your bank being hacked or having a lawsuit filed against you is reputationally negative and can be quantified by the cost of mitigation, the loss of revenue and potential legal liability. However, if the event is less impactful, then measuring reputational risk gets harder.

 

For example, one night, a customer perceived that we closed our branch lobby three minutes early and decided to Tweet about it. Over the course of a half hour, there were more than 20 posts around this small event. Some of the posts were negative; some were positive with other customers coming to our defense. Still, other posts were neutral simply telling the frustrated customer to use the ATM or drive-thru. While the event was negative, the amount of customer support that was shown was likely a net positive. If you were an outsider reading those Tweets, you might conclude that while CenterState might have closed early (or the customer’s watch was off), in general, it was an isolated incident.

 

In similar fashion, one Saturday morning, a non-customer was crunched for time and was trying to cash a CenterState check. They took to Twitter to find a branch and to express their frustration. Luckily we were able to not only provide the needed information but was able to set it up so the branch manager expected them. Anybody monitoring that 10 Tweet interaction could only conclude that we go out of our way to help customers and non-customers alike.

 

Getting Rated

 

Then there are our reviews on places like Glassdoor, Yelp, and others. Here, it should be easy to measure reputational risk since each review comes with a rating. Unfortunately, using ratings are not so clear.

 

Is a rating of three out of five stars positive or negative to your reputation? Is an average Yelp rating with 100 reviews better than a branch with just one positive review?

 

What about banks where the customers don’t care to leave any ratings? 

 

Glassdoor Review:

 

Review Management

 

 

Sample Yelp Review:

Review Management

 

Using Sentiment Analysis

 

Because of the above-mentioned challenges, one framework that we are working with is to grab all mentions of our bank and dump them into a text analytics engine to measure sentiment. Sentiment analysis is the process of computationally identifying and categorizing opinions expressed in a piece of text, in order to determine if the writer's attitude towards a particular topic is positive, negative, or neutral. While most words are neutral, words like “love,” “great” and “happy” are positive while “frustrated,” “angry” or “disappointed” are negative.

 

Of course, the human language is much more nuanced so while we have been working with sentiment analysis for the past two years, it has been only recently having we felt comfortable that we could distinguish “The branch closed early – brilliant” as negative and “CenterState is dope” as positive.

 

If all this sounds complicated, it doesn’t have to be. There are third-party vendors that can do this for you almost instantly for about $100 per month, or banks can also do it themselves for free with several hours worth of effort, following some simple rules.

 

Below is our sentiment analysis for 2016. As you can see about 15% of the third-party discussions about CenterState were clearly positive, while 4% were negative. Of course, about 1% were ambivalent where they expressed both positive and negative sentiment while the bulk of our discussions, some 83%, were neutral.

 

Sentiment Analysis

 

 

For measuring purposes (and board reporting), this could be expressed in a single number in the form of a 3.75:1 positive to negative “Sentiment Ratio.” This can be trended over time so management can better understand if the bank’s service and marketing efforts are creating positive or negative sentiment.

The underlying assumption here is that like movie and product reviews; there is academic evidence to suggest that there is some correlation (about 30 to 45%) between sentiment and revenue (HERE and HERE).

 

Verbal Neutrality

 

The concept of “verbal neutrality” probably merits some discussion within your bank.  For example, when we first tracked consumer sentiment, items like a merger were largely considered positive. Here, it is important to evolve your thinking to not only take into account your bank’s point of view but also the views of the public. Current customers may be concerned with your bank getting distracted by the merger, while customers of the acquiring bank may consider any change to be a net negative. On average, therefore, the net impact of matters like mergers are largely unclear which is why we now classify them as neutral.

 

More importantly, it is important to remember that you are looking to measure opinions that are generated from facts. Opinions are second order results derived from facts. Thus, reporting earnings, no matter how positive, is a sentiment-neutral event. Opinions from those earnings, however, can then be classified as positive or negative.

 

Going On The Offensive – The Jackpot of Unmined Value

 

The above framework allows banks an easy and consistent way to quantify reputational risk. While valuable enough, there is also more to be gained as this analysis can be mined for actionable intelligence.

 

For example, looking further into the data, not only can we tell where the reviews are coming from, but we can also aggregate authors and channels to decipher which writers are more positive and more negative. For those sour on our bank, we can launch a “charm offensive” providing them information that might better inform them or establishing a personal rapport to help establish a better line of communication.  For example, for CenterState, the American Banker staff provides us with the most mentions followed by several papers in Florida. Looking past the obvious outlets, we can also see not only who mentions us on social media, but also their background, gender, education, if they are a parent, marital status, geography, other posts and even friends.

 

From this data, one thing that is surprising is the number of stay-at-home mothers that home school their kids that discuss CenterState. In fact, not only can we ascertain the general sentiment, but the “Power” and “Influence” of that sentiment. By researching each author, we can view their network, the frequency of their posts, the time and day they are most likely to post about CenterState, the ability to influence others and the participating event that generated a post.

 

Thus, we know that if we want to generate more social media sentiment on Facebook regarding our retail banking products, females between the ages of 29 and 35 give us our largest boost and it is usually a result of our bank supporting some cause that they believe in. Here, women are not only more likely to post, but also more likely to re-post, thus resulting in more power and influence per post. Comparatively, the easiest way to get males talking about the retail side of our bank is to generate news around a college alumni affiliation or a sporting event. However, these posts have proven to have less power and influence than their female counterparts.

 

If our goal was to generate positive sentiment about the bank (instead of just our core banking operations), then the fastest way to do that is to release public information that could influence our future earnings such as a merger, new strategic initiative or new product launch. Here, men are three times more likely to talk about our investment potential than women and more apt to re-post through their networks.

 

Conclusion

 

In future posts, we will further explore how using text analytics can drive marketing and how to use analytics around reputational risk to influence sentiment. Until then, if your bank has a need to quantify reputational risk, utilizing sentiment analysis of online text is a proven way to efficiently quantify, monitor and analyze the data. Not only will the effort pay off in managing reputational risk, but it is highly actionable and can directly influence the bottom line.