Arina Tsirkuleva

Data & Tools

• Data: 400,000+ Google Play reviews (2011–2025) · 19 Italian banks (cross-referenced against ECB Significant Institutions list)
• Libraries: google-play-scraper, pandas, matplotlib
• Categories: Large Traditional · Digital-first · Bancassurance · Mid-size Commercial · Consumer Credit · Absorbed

The Rating Gap: Traditional vs Digital

Digital-first banks average 4.32 stars on Google Play. Italy's six large traditional incumbents average 3.42. That gap opened around 2016, when PSD2 forced open banking onto the agenda, and has not closed since: digital-first banks held their scores consistently above 4.0, while large traditional banks moved in the opposite direction. In 2020, COVID pushed millions of first-time users onto traditional banking apps. Scores briefly recovered: 5-star reviews peaked at 53% that year.

The polarity chart makes the structural difference concrete. Among digital-first reviews, 84% are positive and 11% are negative. Among large traditional banks, 34% are negative, i.e., one in three customers who wrote a review was dissatisfied enough to say so. The two absorbed banks in the dataset, UBI Banca and Banca Carige taken over in 2021 and 2022 consequently, averaged 3.92 before disappearing from the market.

The Reply Paradox

The aggregate reply rate across all banks is 43%, hence fewer than half of all reviews receive any response. But there is a wide spectrum of strategies. Mid-size commercial banks (Banca Sella, Credem, ICCREA) have been the most consistent responders, reaching 94% reply rates in 2020 and holding above 90% since. Challenger banks sit at the opposite extreme — near 0% throughout the entire observation period. Large traditional banks improved sharply around 2020, e.g., UniCredit jumped from 17% in 2019 to 95%+ in 2020 in a single year, suggesting a centralized strategic decision rather than organic improvement.

Interesting, that FinecoBank has the highest average score in the dataset with 4.47 stars, yet has never publicly responded to a single review across 15 years and 53,000 ratings. Banca Sella replies to nearly 100% of reviews and scores 4.28. Both seems to be viable strategies. Yet, to understand what works and for which bank we need to the language of the reviews. A preliminary analysis with lagged regression (bank and year fixed effects) shows that a 10 percentage point increase in reply rate is associated with a 0.036 point decrease in average score the following year (β = −0.36, p = 0.06, R² = 0.41). Hence, banks that reply most are managing a high volume of complaints, so the reply rate is a consequence of dissatisfaction, rather than solution.

The breakdown by star rating shows how banks use the reply function: across all categories, 1-star reviews receive higher response rates than 5-star reviews. Review management in Italian banking is complaint handling, not customer engagement. A text analysis of what banks actually write in their replies, and whether content differs between high and low-performing banks is planned as part of the NLP extension of this project.

Data & Tools

• Data: 20,365 central bank speeches from BIS · 130 institutions · 1,003 speakers · 1997–2026
• Embeddings: HuggingFace bge-base-en-v1.5 · 768 dimensions · 101,732 chunks
• Vector store: Pinecone (cloud)
• LLM: Google Gemini 2.5 Flash
• Framework: LlamaIndex · Streamlit

Method

1. Data pipeline: loaded 20,365 speeches, filtered short texts, built institution column.
2. Query routing: temporal, institution, institutional comparative, temporal comparative, full mix.
3. Generation: 5 prompt templates with Chain-of-Thought reasoning.

Example questions

• How did ECB and Fed differ on inflation in 2008 vs 2020?
• What did Draghi say about quantitative easing?
• How has forward guidance evolved since 2008?
• How does the central bank of China differ from the ECB on growth?

Data & Tools

• Data: 50,000+ customer reviews (2010–2022) scraped from Google Play Store for banks under direct ECB supervision.
• Libraries: google-play-scraper, langdetect, spaCy, NLTK.
• Visualization: Tableau

Method

1. Scraping & Language Detection: Collected reviews per bank app, detected the dominant review language. Due to the unavailability of spaCy pipelines for several languages (in particular, Latvian, Finnish, and Estonian), I excluded these banks from the analysis. For Greek banks, Greek and English showed up almost equally, so I kept both in the analysis.
2. Preprocessing: Cleaned text, removed stopwords + generic words ("bank", "application").
3. NLP Analysis:
   • Used spaCy pipelines for nouns (issues/features) and adjectives (sentiment/attitudes).
   • Counted most frequent 10 nouns + 10 adjectives per star rating (1–5).
4. Visualization: Built an interactive Tableau dashboard showing:
   • Sentiment evolution 2010–2022.
   • Top praised vs criticized app aspects.

Findings

• Praised: update, ease in making payments, ease of use, intuitive design.
• Criticized: bugs in updates, security, connection, login issues, activation problem, speed, lack of personalization.

Impact

The project demonstrates how AI/NLP on unstructured customer data can support:

• Banks in prioritizing digital strategy investments.
• Regulators in understanding evolving customer expectations in financial services.

The European Central Bank (ECB) was originally designed as a technocratic institution. The Maastricht Treaty set out the ECB’s formal structure and mandate, yet the ways in which these are exercised have been gradually crafted by the institution’s leaders. Members of the Executive Board not only manage monetary policy but also define the ECB’s public voice. Through their speeches and statements, they construct and project the institution’s identity.

I examine how that identity has evolved by focusing on the modality of ECB executives’ communication. In linguistics, modality refers to the speaker’s attitude toward reality or truth. There are several ways to classify modality, but two categories are especially relevant here:

• Epistemic modality: expresses the speaker’s stance toward knowledge, belief, or certainty in a proposition.
• Deontic modality: reflects how the world should be according to norms or expectations.

Data & Tools

• Data: ECB Speeches Dataset (1999–2021).
• Libraries: NLTK, spaCy (for preprocessing and modality parsing).
• Custom Resources: Manually built dictionary of modality markers (epistemic: may, might, will; deontic: must, should, we commit to...).

Method

1. Preprocessing: Tokenization, lemmatization, stopword removal across multiple languages.
2. Modality Parsing:
   • Identified epistemic modality (beliefs, predictions, probabilities).
   • Identified deontic modality (norms, commitments, obligations).
   • Constructed a custom dictionary of modal expressions and phrases, mapped to spaCy parsing outputs.
3. Temporal Analysis: Tracked shifts in modality use over 22 years of ECB communications.

Findings (work in progress)

• Crisis periods (2008–2012): Higher frequency of deontic modality (commitments, obligations), signaling strong intervention stance.
• Post-crisis and integration phases: Growing use of epistemic modality, emphasizing probabilities, scenarios, and market expectations.
• The ECB’s communication style has contributed to its institutional empowerment, moving from “technocratic” to “performative actor” in European integration.

Impact

The project demonstrates how AI/NLP can systematically analyze performativity in policy communication, offering insights into:

• How central banks shape market expectations beyond quantitative measures.
• How linguistic strategies reflect evolving institutional power (Brunnermeier et al., 2016; Chang, 2015; Tortola, 2020).
• Potential early-warning indicators of market-moving communication.

Data & Tools

• Data: Central bank speeches from the Bank for International Settlements (2007–2021), scraped via Selenium.
• Libraries: spaCy (NER & preprocessing), NetworkX (graph analysis), Louvain method (community detection), Pyvis (interactive visualization).

Method

1. Named-Entity Recognition (NER): Extracted organizations and individuals mentioned in central bank speeches.
2. Cleaning & Normalization:
   • Removed self-references (e.g., ECB mentioning "ECB").
   • Merged coreferences (e.g., "Fed", "Federal Reserve System", "US Federal Reserve"): automatically using ECB abbreviations glossary and manually.
   • Applied a frequency threshold (≥10 mentions).
3. Network Construction:
   • Built directed source–target networks: speech source → entity/person mentioned.
   • Separate networks for organizations and individuals.
4. Community Detection: Applied Louvain method to identify clusters of closely connected institutions and policymakers.
5. Visualization: Interactive Pyvis graphs (suggested ForceAtlas2 solver for clarity).

Findings

• Organizational network:
  • ECB, IMF, and US Federal Reserve emerge as central hubs.
  • Regional clusters (e.g., European vs. Asia-Pacific banks) highlight patterns of co-mention and influence.
• Individual network:
  • Prominent leaders (Draghi, Bernanke, Lagarde) occupy bridging positions across clusters, illustrating how personalities drive institutional connectivity.

Impact

This project demonstrates how AI-driven network analysis of communication can uncover hidden structures of influence in global finance. It provides tools for:

• Mapping policy interdependence across central banks.
• Identifying influential individuals and organizations in crisis vs. normal times.
• Offering a new perspective for both researchers and policymakers on coordination in global financial governance.

Data & Tools

• Data: Introductions to annual reports of 11 NCBs (1999–2020), manually collected from official sites.
• Libraries/Methods: Topic modeling (25 pre-defined topics), Hellinger distance for text similarity, Tableau for visualization.

Method

1. Corpus Construction: Compiled introductions from ECB and NCB annual reports. (Portugal excluded due to missing data).
2. Topic Modeling: Represented each text as a distribution of 25 predefined topics (e.g., inflation, financial stability, growth).
3. Distance Calculation: Applied Hellinger distance to measure the similarity between ECB and NCB communication in each year.
4. Visualization: Built an interactive Tableau dashboard to show how distances evolved by country and over time.

Findings

• High alignment in crisis periods: ECB and NCBs converged during 2008–2012, reflecting coordinated responses.
• Divergences post-crisis: Some NCBs (e.g., Greece) displayed greater distance from ECB discourse, linked to national economic challenges.
• Long-run trend: Overall, ECB and NCB communications became more consistent, though national differences persist.

Impact

The project illustrates how text-based distance metrics can measure institutional alignment within the Eurosystem. Insights are relevant for:

• Assessing the effectiveness of common monetary policy.
• Detecting gaps between supranational and national communication that may affect credibility.
• Supporting policymakers in ensuring coherent messaging across the euro area.

Data & Tools

• Data: Introductions to annual reports of 11 NCBs and the ECB (1999–2020), manually collected from official websites.
• Libraries: NLTK & spaCy (preprocessing), Gensim (phrase detection & topic modeling).
• Visualization: Tableau dashboard.

Method

1. Corpus Construction: Collected introductions of annual reports for all euro area founding members (plus Greece from 2001). Portugal excluded due to missing data.
2. Preprocessing: Tokenization, lemmatization, and stopword removal; merged smaller paragraphs into comparable text chunks.
3. Topic Modeling: Trained an LDA model (25 topics) on the full corpus to capture recurring themes across time and institutions.
4. Visualization: Built interactive Tableau dashboards showing:
   • Average attention to topics across the Eurosystem.
   • Topic dynamics for each central bank over time.

Findings

• Crisis years (2008–2012): Sharp rise in attention to financial stability, sovereign debt, and market support.
• Post-crisis: Increasing emphasis on growth, and structural reforms.
• ECB vs NCBs: While overall trends converged, national banks highlighted local issues (e.g., employment in Italy, competitiveness in Spain, fiscal policy in Greece).

Impact

This project provides a systematic mapping of the building blocks of Eurosystem communication. It highlights how:

• Central banks shift their focus in response to global and national shocks.
• Supranational and national institutions balance common messaging with local priorities.
• Text mining can serve as a quantitative lens for policy communication analysis.