Case Study
№ 01

Fintech Sentiment Intelligence Analysis.

AI-powered analysis of 10,386 fintech app reviews with custom severity detection: separating emotional noise from real operational impact.

10,386 Reviews4 AppsVADER + NegationCrisis LexiconSeverity 1:5Streamlit Dashboard

Artifacts

Cleaned dataset (10,386 rows), preprocessing pipeline, negation-aware VADER wrapper, 50+ keyword crisis lexicon, 1:5 severity scorer, and a 4-page interactive Streamlit dashboard with app / sentiment / severity filters and competitive benchmarking.

01Business Problem

Sentiment tools score tone, not impact.

Standard sentiment tools (VADER, star ratings) miss high-severity complaints in fintech reviews. They score emotional tone, not operational impact.

A calm complaint about "account locked for 3 weeks" reads as NEUTRAL. An angry complaint about "ugly UI colors" reads as NEGATIVE. For a fintech company, missing the account lockout is a product-risk failure, not a sentiment error.

The goal: build a system that separates emotional noise from real operational impact, so engineering and support teams can prioritize what actually hurts users.

02What I Built

A 6-stage severity detection pipeline.

Step 1

Data Collection

Scraped 10,400 reviews from the Google Play Store, 2,600 per app across Cash App, Chime, PayPal, and Venmo. Cleaned to 10,386 reviews after removing 14 duplicates.
Step 2

VADER Baseline

Ran standard VADER sentiment to set a floor. Result: a 26% miss rate on 1:2★ reviews (scored as neutral or positive), which motivated everything that followed.
Step 3

Error Analysis

Hand-labeled 200 reviews and diffed them against VADER output. Two dominant failure modes emerged: negation errors and missing crisis vocabulary.
Step 4

Negation Handling

Pre-processes phrases like "not good" so the sentiment correctly flips to negative. Negation errors dropped from 38.4% to 36%.
Step 5

Crisis Keyword Lexicon

Built 50+ fintech-specific keywords across account access (frozen, locked), fraud (hacked, scam) and money issues (missing funds, stuck transfer). Auto-flags reviews for a severity boost.
Step 6

Severity Scoring (1:5)

Combines sentiment intensity, crisis keywords and star rating into a single 1:5 score. Ranks complaints by business impact instead of emotional tone.
Step 7

Interactive Dashboard

4-page Streamlit dashboard with real-time filtering by app, sentiment and severity, plus competitive benchmarking charts for side-by-side app comparison.
03Results

Severity detection improved 87%.

Model performance measured against 200 hand-labeled reviews. All competitive numbers are computed on the full 10,386-review corpus.

Severity macro-F1
0.16 : 0.30
+87% over VADER baseline
High-severity F1
0.55
up from 0.47
Sentiment accuracy
67%
on 200 hand-labeled reviews
Negative share
30.6%
of all 10,386 reviews (rating ≤ 2)

Crisis Pattern Mix

Where the model surfaces high-severity complaints once the emotional-tone bias is removed.

Account access issues
35%
Money / payment problems
28%
Fraud & security
18%

Competitive Ratings

AppRatingReviews
Cash App4.07 ★2,600
Chime3.89 ★2,595
PayPal3.43 ★2,594
Venmo3.31 ★2,597

Where VADER underweights complaints

Fig. 01Volume of high-severity complaints VADER misclassifies, and the miss rate per severity band.

Complaint volume by app × severity

Fig. 02Heatmap of complaint counts across apps and severity bands, from Low to Critical.

App-level severity benchmark

Fig. 03Per-app misclassification rate, average severity, and the mix of Critical + High complaints.

Complaint themes per app

Fig. 04Grouped complaint themes surfaced per app, from account issues to support friction.

Average severity per theme

Fig. 05Which complaint themes carry the highest average severity score across the corpus.
04Business Impact & Actions

What decisions change once severity is measurable.

The point of the model is not a better score. It is a different set of actions across product, support, and marketing.

For Product Teams
  • Prioritize engineering fixes by severity score, not review volume. Route severity 4:5 issues into sprint planning automatically.
  • Track which complaint categories correlate with churn and focus roadmap on account lockout and stuck transfer fixes first.
  • Benchmark against competitors' pain points and use it as market-positioning input.
For Customer Support
  • Auto-flag severity 4:5 cases for immediate escalation. Severity-based routing replaces manual triage.
  • Surface complaints missed by keyword filters so support reviews flagged hidden negatives weekly.
  • Cut response time on money and fraud issues with specialized queues for crisis-keyword reviews.
For Marketing
  • See what competitors' users complain about most, then craft messaging that addresses those pain points directly.
  • Identify your product's relative strengths and turn a lower account-lockout rate into a differentiator.
  • Feed verified competitor complaints into positioning and creative reviews.
What decisions change
  • Support tickets routed by severity, not arrival order.
  • Engineering sprints prioritize high-severity complaint categories.
  • Marketing messaging addresses verified competitor pain points.
  • Retention team intervenes on accounts with severity 4:5 complaints.
Revenue Impact
  • Reduce churn by catching high-severity issues faster.
  • Improve retention by fixing complaints that drive account closures.
  • Enable better competitive positioning through pain-point intelligence.
05What Worked / What Failed

Honest results, honest limitations.

What Worked
  • Crisis keyword lexicon caught complaints VADER missed.
  • Negation handling reduced errors by roughly 6 percentage points.
  • Severity scoring cleanly separated tone from impact.
  • Competitive benchmarking exposed app-specific weaknesses.
What Failed / Limitations
  • Sentiment accuracy still only 67%. Target was 75%.
  • Sarcasm: "Great, now my account is locked" still scored positive.
  • Slang and informal language like "my card got cooked" not flagged initially.
  • Very short reviews like "Trash" do not cluster meaningfully for topic analysis.
Reality Check

VADER alone missed 26% of 1:2★ reviews, scoring them as neutral.

Star ratings are misleading: many 3:4★ reviews contain severe complaints buried inside otherwise calm text.

First-pass keyword scoring failed on non-standard phrasing, which is why the lexicon and negation layer both exist.

06What I'd Improve Next

Where this pipeline goes next.

Short-term / 1:3 months
  • Fine-tune BERT to push sentiment accuracy from 67% toward 80%+.
  • Expand crisis lexicon with slang (cooked, chalked, finessed).
  • Add a sarcasm detection layer.
  • Build a severity trend dashboard with week-over-week deltas.
Medium-term / 3:6 months
  • Real-time monitoring with automated daily scraping and alerts.
  • Multi-language support (Spanish, Portuguese for LatAm fintech).
  • Supervised severity classifier trained on 2,000+ labeled reviews.
Long-term / 6:12 months
  • Causal modeling: which complaint categories actually drive churn.
  • API deployment for production integration.
  • Wire into support ticketing for automatic severity-based routing.
07Why This Matters

Reviews are a strategic signal, not noise.

What Actually Drives Churn
  • Locked accounts and access failures.
  • Missing money and stuck transfers.
  • Failed payments and deposit delays.
  • Fraud scares and security breaches.
  • Slow support response during crises.
The Strategic Value
  • Risk teams see crises earlier.
  • Support teams escalate the right cases.
  • Product teams fix issues that drive churn.
  • Marketing teams position against competitor weaknesses.
  • Leadership gets a real-time view of operational pain points.
Why This Project Matters to the Business

Fintech companies do not lose customers because of bad UI colors. They lose them because of operational failures: locked accounts, missing money, failed transfers, fraud scares, and slow support.

These issues create churn, regulatory exposure, reputational damage, social media blowups, support overload, and revenue loss.

Severity detection turns customer reviews from noise into a strategic signal. It connects data work directly to revenue, risk, and leadership priorities.