Leveraging social media to address customer needs: Investigating tweets to Revolut

Photo by Viktor Forgacs on Unsplash

Introduction

Twitter is a powerful tool for businesses, allowing them to connect directly with their customers and reach a broad and engaged audience. For businesses with a large and engaged Twitter following, this platform can become a valuable source of customer insights on how to improve products and services.

Given this context, for my fourth project in the 12-week Metis Data Science Bootcamp, I analyzed customer tweets to Revolut to identify opportunities to address customer needs.

Revolut background

Revolut is a UK-based financial technology company (fintech) that launched in 2015, initially offering money transfer and exchange. Since then, they have grown their product offering and expanded into countries across Europe, as well as the United States, Canada, Australia, Singapore and Japan.

In February 2020, the company raised £387 million at a valuation of £4.2 billion, becoming the most highly-valued fintech in the UK. By December 2020, Revolut’s customer base was estimated at 14.3 million.

Methods

Both quantitative and qualitative methods were used as part of the analysis. These are depicted in the figure below.

Methods

Data

The analysis included 97,000 customer tweets from January to December 2020. These tweets were scraped using Twint, an advanced Twitter scraping tool.

As exhibited below, there is a consistent number of tweets overtime, despite some minor fluctuations. There is however, a small jump in the volume of tweets in March as well as a larger jump in June. These movements are likely linked to two big events for Revolut; in March, the company launched in the United States¹, while in June a new version of the app was released.²

Number of tweets over the12 months of analysis

While the volume of tweets is noticeably smaller in January, this appears to be linked to seasonality, rather than any specific event.

Findings

Understanding the main tweet topics

Topic modeling using non-negative matrix factorization (NMF) revealed four major tweet topics and the top 6 words within each topic. These are depicted in the figure below.

Tweet topics

General account queries: These are general queries related to opening, accessing and using a Revolut account.

App queries: Includes a range of app-related topics, including updates, bugs, and logging in and using the app.

Fintech Innovation: Relates to Revolut as a fintech with innovative product offerings. ‘Child’ appears because in 2020, Revolut launched the JuniorApp, targeting 7–17 year olds.

Account access: This is a very specific account-related topic, concerning an issue with accounts being frozen for review. This topic generally includes complaints that customers are not able to access their money to pay their bills, sometimes for a period of months.

Monitoring tweet emotions

It is also important to understand the sentiment of tweets to keep in touch with the company’s audience and monitor what is being said about the brand online. This is particularly important because negative social mentions can be harmful to the brand image. To enable this analysis, I used Vader, which has been “specifically attuned to sentiments expressed in social media.”

The figure below depicts the average sentiment per topic over the 12 months of analysis. Integrating sentiment analysis reveals that tweets are generally positive, although close to neutral.

Sentiment by topic over time

As you can see in the image, the App queries and Account queries topics are generally neutral and appear to fluctuate a bit less over time. Therefore, zooming in on Fintech innovations and Account queries, we can clearly depict how peaks and dips appear to correspond to important moments for Revolut in 2020.³ ⁴ ⁵ This is reflected in the figure below:

Annotated sentiment by topic over time

Identifying tweet intention

I also wanted to understand the intention behind tweets, in terms of whether the customer is submitting a compliment, query, complaint or other. To do this, I developed an intent classifier using XGBoost. I started by hand labeling the intent of approximately 1,200 tweets and then predicting on a subset of 300 tweets. I then reviewed and corrected predictions where necessary, and added those data points to the training data. I conducted this process iteratively until I had a training set of approximately 3,600 tweets.

The figure below depicts the performance of the classifier, in terms of the percentage of each class that the model correctly classified.

Percent of correct predictions by intent

As shown in the figure, the model performed well in classifying the larger classes of Complaints and Queries. In contrast, it performed less well in predicting the smaller classes, particularly the Other class. This is a common challenge with class imbalance; however, given the number of classes, this was considered good performance overall.

Incorporating customer intent with sentiment enables insight into how to address different categories of customer tweets. The figure below depicts the distribution of sentiment within the larger classes, Complaints and Queries.

Share of sentiment by intent: Company and Queries

One interesting observation from the figure is that there are complaints that have been classified as having a positive sentiment. Digging into these a bit further, this sheds light on one of the challenges with sentiment analysis — correctly identifying sarcasm as having a negative sentiment.⁶ For example:

“Hello, it would be kind if I can talk with someone to unlock my account. Except Liza who is acting like a robot. Thanks.”

Reading the above tweet, we can identify that the customer is insulting the service representative with whom they have already engaged. However, the presence of ‘kind’ and ‘Thanks’ has led the model to classify the tweet as having a positive sentiment.

While it is unsurprising that customer complaints are predominantly negative, these findings still provide important insights. Customers who are submitting a complaint and have a strongly negative sentiment can become harmful to the brand image online as they may be inclined to attack and tarnish the brand image. Therefore, it is important to address these customers appropriately.

While it is promising to see that the majority of Queries are positive, a large share is still negative. It is similarly important for Revolut to manage these appropriately before customers become frustrated and enter into the Complaint territory, which can be more difficult to resolve.

Insights

Together, these findings provide helpful insights for Revolut, as well as other companies on social media.

When there is a new version release, expect an increase in the volume of tweets and a reduction in sentiment overall as there are often small bugs that crop up.

  • Consider the use of message pop-ups to encourage use of in-app support; it is important to remind users that in-app support is available, to keep Twitter as a very last resort for requests for support.
  • Increase support staff during this period to timeously and appropriately address the volume of queries and complaints.
  • Analysis of in-app support would be helpful as a large volume of tweets complained that in-app support either was not working or that they could not locate it within the app. Investigate opportunities to improve in-app support for usability.

It is important to appropriately address complaints online.

  • A large volume of tweets are complaints, which can be potentially harmful to the brand’s image. Begin by filtering out the bots, and address the remaining complaints so that they are not provoked into becoming anti Revolut.
  • Focus support on customers classified as submitting queries, prioritizing those with a negative or neutral sentiment first. Again, it is important to address these customers before they become increasingly frustrated and become Complaints.

As the company continues to grow, so too will the volume of communication with customers. Therefore, it is important to address these opportunities now.

[1] R Dillet, Revolut launches its neobank in the US (2020), TechCrunch.

[2] N Storonsky, Introducing Revolut 7.0: One app to manage all things money (2020), Revolut Blog.

[3] K Makortoff, Digital bank Revolut becomes UK’s most valuable fintech startup (2020), The Guardian.

[4] R Hinchliffe, Revolut accused of coercing 50 workers to quit their jobs (2020), Fintech Futures.

[5] E McGrath, You can now buy silver on Revolut (2020), Revolut Blog.

[6] It is important to note that although useful, sentiment analysis is not without its shortcomings. This is particularly the case with social media text, which is often full of colloquialisms and emojis, and can alter the meaning of text.

always a student

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store