Text Analytics on the Consumer Complaints Against Chime Bank
A report on analyzing consumer complaint narratives related to Chime Bank using SAS platform.
Background
As part of my job as a technical intern at SAS this past summer, I used SAS Visual Text Analytics (VTA) and SAS Visual Analytics (VA) to uncover some emerging trends among the complaints related to the Consumer Financial Protection Bureau (CFPB). On July 6th, ProPublica published an article explaining that Chime bank is receiving complaints from customers who are unable to access their funds. According to the business, there has been an “exceptional spike” in fraudulent deposits, but that is little consolation for customers caught in the middle of the chaos.
Chime is the largest of a rising type of financial technology businesses known as “neobanks,” which cater to low- to moderate-income customers who are underserved by traditional banks. A neobank, in particular, is a type of direct bank that operates entirely online and does not have traditional branch networks. They had a great year the prior year. Chime is a good example of this since it grew from zero to eight million users between 2013 and 2020, then soared to twelve million in the last year. The funds for many of the new accounts came from government stimulus payments.
However, there has been a high volume of complaints to the CFPB related to customers being unable to access their money or accounts, and that Chime is slow to resolve issues, among other things. Meanwhile, several Facebook groups have sprung up with titles like “Chime Bank has FAILED” and “Chime Thieves.” They have a total of 700 members. Alex Horowitz, senior research officer for the consumer finance project at the Pew Charitable Trusts states “since Chime is not a bank, that leaves it in a regulatory no man’s land.” As a result, the rules and jurisdiction are only murky at best. Regardless, state officials have taken note of Chime. In late 2019, the California Department of Financial Protection and Innovation received complaints that customers were unable to access their accounts and pay their bills due to a Chime system failure. After reading this story, I decided to conduct additional research to see if there were any more complaints and what else I could uncover.
I’ll go into more detail about how I used VTA to evaluate the consumer narrative dataset in the next section. Skip to “Exploration with SAS Visual Analytics” if you simply want to know about the results.
Preparing Data
In this project, I analyzed the Consumer Complaint dataset (between the years 2018–2021) obtained from the Consumer Financial Protection Bureau website. I brought that data into SAS Data Studio in which I created an ID column with unique values and in SAS Model studio I began to run some visual text analytics.
For variables such as “State”, “Issue”, and “Product” I assigned them the role of a categorical variable. For “Consumer Narrative” I assigned it to be a text variable as that would be the variable that I would be creating concepts and categories for. Through Model Studio, I was able to create a VTA pipeline, which is a process flow diagram whose nodes represent tasks in the text analytics process. The nodes I focused on during this project were concepts, sentiment, and categories.
A concept is a data element or pattern — such as named entities or fact relationships — that you wish to extract from a larger text field because they match a specific context. In VTA, I wrote rules recognizing concepts that were important to the context of the fraud committed by Chime Bank. The four main concepts I created were: Cannot Access, Closed Account, Stolen Money, and Unauthorized Transfers. These concepts were created because as I used visual text analytics to interactively explore the narratives, I found that people tended to have problems with not being able to access their account, Chime would close their account without an explanation, customers couldn’t access their account funds, or they saw unauthorized money transfers show up on their bank statements. These concept rules use language interpretation for textual information (LITI) syntax, which includes boolean and distance operators. For example, in the Unauthorized Transfers concept, my rule was Concept_Rule: (Sent, “_c{transferred@}”, (OR, “money@”, “knowledge@”)). This rule looks for any synonyms of transferred in the same sentence with the synonym of money or knowledge as a way to fetch all the consumer narratives in which the consumer complains of having their money transferred from their account without their knowledge. In the concept node, I also included VTA’s predefined concepts that include the following: Person Names, Location Names, Organization Names, Dates, Times, Currency Amounts, and Percentage Amounts. Overall, concepts are important because they influence the way in which text is parsed as we are pulling out specific pieces of information.
In my VTA pipeline, I also created custom categories through the process of tagging documents as belonging to a specific category using linguistic rules. For this project, I created categories such as Managing an Account, Unauthorized Transactions, Fraud or Scam, and Closing Your Account and SAS VTA generated other linguistic rules using supervised learning on how the document issues were manually tagged. Categories are rules generated based on boolean logic for the presence of specific terms and phrases; the code uses OR, AND, and NOT commands to dictate what keywords will be looked for in the consumer narratives. As seen in Figure 1, I am looking for keywords like “access”, “lost”, “date”, and “deposits” to find narratives of consumers complaining about not being able to access their Chime Bank account. In the categories tab, we can see all the consumer narratives that are matched and their sentiment.
After creating all my concepts and categories, I ran the VTA pipeline to be able to download the score code for categories, concepts, and sentiment from VTA results. The score code was then entered into a code window in SAS Studio and then brought into Visual Analytics to allow exploration and visualization of the data. Since I had score code for categories, concepts, and sentiment, I had to left join the concepts and categories in SAS Studio and save the result as an in-memory table. Then, in SAS Visual Analytics, I could left join the file that had the categories and concepts combined with the sentiment score code. This allowed me to have a file that would have the information of all three nodes in one place so I could visualize all that information.
Exploration in SAS Visual Analytics
In SAS Visual Analytics I choose to have the following sections to analyze the data: Background, Categories by State, and Concepts by State. However, before diving into any individual section, I had to create hierarchical data items to be able to visualize the relationships between the categories I created and the matched text that was fetched from the consumer narratives.
In the background section, the title and subheading give the audience a quick overview of what this report will be about. The word cloud shows the top 10 issues that were prominent in the consumer narratives. We can see that consumers tend to complain about unauthorized transactions, Chime Bank closing their account, not being able to retrieve their funds from their closed account, and problems accessing their account. To find out where, when, and why these complaints were specifically made, we will be using our categories and concepts sections.
One of the categories that has the most amount of consumer narratives (246) related to it was “Closing an account.” Inside this category, “closed” was the matched text with the highest negative sentiment score and most consumer narratives. On the right of Figure 2, we can see that this issue was most prominent in Arizona, Texas, Arkansas, and Massachusetts. This visualization is possible as SAS VA allows linking between objects.
By double-clicking on this matched text, we can see the specific consumer narratives and the frequency at which they were received based on the dates. In Figure 3, the first consumer narrative states “a couple months ago Chime closed my account without notice… they’ve blocked my number from calling & not responding to my emails.” This complaint seems to have been received twice; once on 01/21/2021 and on 03/25/2021.
From Figure 4, by clicking on the category of “Unauthorized transactions or other transaction problem” we can see this problem is prominent in Georgia, but also apparent in California, Oregon, Minnesota, Tennessee, Florida, New Hampshire, and Louisiana. One of the consumer narratives within this category states “Chime bank is a deeply negative crooked scam company… I had an account 4 months, noticed an unauthorized transaction… they basically can close your account for no reason WITHOUT A REFUND.”
In the concepts section, I created concepts that were similar to the categories to see if I could fetch more interesting consumer narratives. However, I also made a concept called “Stolen Money” in which consumers complain of having Chime Bank stealing their account funds with a high negative sentiment score. This concept is notable in Texas and Virginia as seen in Figure 5. One of the consumer narratives states “Chime Bank stole almost {$4000.00} from me. They put a false deposit on my account. And they closed my account due to that.”
Data Findings
Overall, almost all 50 states have been affected by consumers not being able to access their account, having their account closed, seeing unauthorized transactions, or having their funds stolen by the bank from their account. This neobank is using the argument that they are taking these actions on these accounts to combat fraud. Yet, despite many of these consumer complaints being resolved as “Closed with an explanation,” many of these consumers are confused by the actions the bank has taken on their accounts and are being affected by damaging financial conditions. Chime consumers who cannot access their funds have been unable to pay their bills, were subject to late fees, and are at risk of facing home foreclosure.
Using a visual text analytics framework like this is interesting — but it could also assist regulators to investigate consumer complaints faster and more effectively. This should therefore help to strengthen consumer protection and clamp down on unfair, misleading, or abusive financial market activities. Financial providers and other companies could also use this kind of framework to dig deeper into their own complaints data. They could therefore identify problems and issues affecting their customers, and respond proactively, without waiting for regulatory findings. Indeed, rapid and well-directed action in response to customer complaints could, ultimately, mean that they are not subject to regulatory investigations.
It is important to stress that this kind of text mining framework is not intended to replace human analysts in any way. However, it could be a useful aid to those analysts in identifying particular problems or geographic areas that have experienced issues. Ultimately, these capabilities enable regulators like the CFPB to better protect consumers from risky financial products and services. In the end, the CFPB’s mission is to protect consumers from risk from financial products and services.
