AI Reads a Million App Reviews to Reveal Why Consumers Distrust Open Banking

Open banking was supposed to transform finance by handing consumers control of their own data, letting apps aggregate accounts, compare products, and trigger payments across institutions with a single tap. Regulators, particularly in Europe under the revised Payment Services Directive, have pushed hard for this data-sharing ecosystem, and FinTech developers have responded with a wave of open banking applications. Yet adoption has stubbornly lagged behind the regulatory ambition. A new study published in Neural Computing and Applications offers one of the most detailed explanations yet for that gap, and it comes not from surveys or focus groups but from the unfiltered words of more than a million app users.

Researchers led by Mohamed M. Mostafa of the Institute of Business Administration in Karachi, together with colleagues at the University of Manchester, Lamar University, and Texas Southern University, assembled a corpus of over one million user-generated reviews of open banking applications drawn from the Google Play Store. Rather than reading a sample and extrapolating, the team applied natural language processing at full scale, using two complementary computational techniques: structural topic modeling and sentiment analysis. The goal was to map what consumers actually talk about when they review these apps, how those themes relate to one another, and how strongly users feel about each of them.

Structural topic modeling is a statistical framework that treats every document, in this case every review, as a mixture of latent themes. The method, developed by political scientists and statisticians including Margaret Roberts, Brandon Stewart, and Dustin Tingley, improves on the classic Latent Dirichlet Allocation approach by allowing researchers to correlate topics with each other and with document-level metadata. The team also ran LDA, the workhorse algorithm of topic modeling, which assumes each document draws its words from a probability distribution over topics, and each topic a distribution over words. By comparing the two models and evaluating topic coherence, the researchers could identify dominant themes in the review corpus with statistical confidence rather than intuition.

The analysis surfaced a set of recurring topics that will feel familiar to anyone who has abandoned a banking app in frustration: customer service, data privacy, and digital support ranked among the most prominent. But the study went beyond simply listing themes. By mapping the inter-relationships between topics, the researchers showed how concerns cluster together, revealing the structure of consumer perception rather than just its surface. That structural view matters because it suggests that fixing one complaint in isolation may not resolve the underlying dissatisfaction if it is tightly linked to other grievances.

The sentiment analysis operated at multiple levels of granularity. At the coarsest level, the team measured overall polarity, classifying reviews as positive, negative, or neutral. They then drilled down to specific emotional tones, drawing on a tradition of emotion mining that includes lexicon-based methods such as SentiWordNet and the crowdsourced word-emotion association lexicon developed by Saif Mohammad and Peter Turney, as well as the psychoevolutionary theory of emotion proposed by Robert Plutchik. The result is a layered picture: two reviews may both be negative overall, but one may express anger about a failed transaction while the other expresses anxiety about data misuse, and those distinctions carry very different implications for how a provider should respond.

The technical pipeline also had to contend with the messy reality of app-store language. Reviews are short, informal, saturated with emojis, and often sarcastic, all of which challenge standard text-mining tools. The researchers’ decision to combine topic modeling with fine-grained emotion analysis reflects a broader trend in computational linguistics, where single-method studies are increasingly seen as insufficient for capturing the complexity of user-generated content. Prior work on mobile banking apps in countries from Morocco to Saudi Arabia and Turkey has used similar techniques, but the sheer scale of this corpus, over a million reviews, makes the new study the first of its kind for open banking applications specifically.

The findings arrive at a critical moment. Open banking’s promise depends on consumer consent: users must agree to share their financial data with third parties before the ecosystem can function at all. The study’s authors frame persistent issues of trust, data privacy, usability, and system reliability as the principal brakes on adoption, and their empirical results give those abstract categories concrete texture. When a million users write about privacy, they are not voicing a vague philosophical worry; they are describing specific experiences with specific apps, and the language they use reveals which aspects of data handling provoke the strongest emotional reactions.

For financial institutions and FinTech developers, the study offers a scalable framework for listening to customers continuously rather than episodically. Traditional satisfaction surveys capture a snapshot of opinion from a self-selected sample; app-store reviews capture a running stream of feedback from the people who actually use the product. The authors position this kind of user-centered data analytics as a strategic tool for enhancing engagement, mitigating user churn, and fostering adoption through transparency and targeted service design. In other words, the same reviews that public researchers mine for insight are, in principle, a free and constantly updating diagnostic instrument for the industry itself.

Regulators have reasons to pay attention too. The study explicitly aims to bridge the gap between regulatory ambition and consumer perception, and its methods could be adapted to monitor how sentiment shifts as new rules take effect. The literature on open banking spans questions of competition, credit access, and financial inclusion, from research on how customer data access enables FinTech entry to analyses of open banking’s impact on lending in emerging economies. Understanding the emotional register of consumer complaints adds a missing behavioral dimension to those policy debates, one that could help explain why legally mandated data sharing has not automatically translated into mass consumer uptake.

The research also contributes to a growing body of work that integrates computational linguistics with behavioral finance theory, treating text as data in the sense championed by Justin Grimmer and Brandon Stewart. That methodological lineage comes with caveats that the authors acknowledge through their careful model evaluation: topic models are sensitive to preprocessing choices, and coherence metrics must be interpreted with care. Still, the study demonstrates that when these tools are applied rigorously to a large, publicly available corpus, they can turn the noise of a million app reviews into a structured, actionable map of consumer sentiment. As open banking matures, the voice of the user, captured at scale and decoded by machines, may prove to be the most important dataset in the room. The data underlying the analysis is publicly available via the Google Play Store API, allowing other researchers to replicate and extend the approach.

Subject of Research: Consumer perceptions of open banking applications analyzed through topic modeling and sentiment analysis of user reviews

Article Title: Analyzing consumer reviews on open banking apps: insights from topic modeling and sentiment analysis

Article References: Mostafa, M. M., Feizollah, A., Sargsyan, G., & Shetty, S. (2026). Analyzing consumer reviews on open banking apps: insights from topic modeling and sentiment analysis. Neural Computing and Applications, 38(17), Article 727. https://doi.org/10.1007/s00521-026-12402-7

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12402-7

Keywords: open banking, topic modeling, sentiment analysis, natural language processing, structural topic model, Latent Dirichlet Allocation, FinTech, data privacy, consumer trust, mobile banking apps, customer satisfaction, text mining