Back to Projects
Financial Regulation / NLP

G-SIB Disclosure Analysis Prototype

An NLP prototype that turns public disclosures from Global Systemically Important Banks into comparable sentiment and risk signals across banks and quarters.

Type
Cambridge employer project (team)
My Role
Designed and built the core pipeline
Data
81 public reports and transcripts, Q1 2023 - Q1 2025
Status
Prototype

The Challenge

FinBERT and VADER scored 17,292 passages from 81 bank reports and earnings-call transcripts, showing which bank's disclosures carried the most negative language and which risk themes drove it.

Supervisors and analysts face more published material from systemically important banks than they can read consistently: quarterly reports, earnings-call transcripts and Q&A sessions. Signals about credit quality, funding and legal exposure are spread across hundreds of pages per quarter.

The brief was to prototype how NLP could make these disclosures comparable across banks and over time, using only public documents.

Approach

01
Document collection and processing
81 public documents for three G-SIBs, Q1 2023 to Q1 2025 (27 per bank): earnings reports and earnings-call / Q&A transcripts. Text extracted from PDF and Word files and split into chunks, with bank, quarter and document metadata kept for traceability.
02
Sentiment scoring
Every chunk scored with FinBERT (finance-specific) and VADER (lexicon-based), then aggregated by bank and quarter.
03
Risk themes and metrics
Negative passages mapped to a keyword-based risk taxonomy (credit, market, liquidity, operational, legal and more). Financial metrics such as capital ratios and net interest margin extracted with regular expressions.
04
Summaries and reporting
GPT-2 document insights and BART-based summary aggregation, with HTML reports and a dashboard image for review. A teammate built a separate LLM summarisation module; the wider team contributed brainstorming and validation.

Results

81
Public reports and transcripts across 3 G-SIBs
17,292
Text chunks scored with FinBERT and VADER
15.8%
Of chunks classified negative by FinBERT

FinBERT classified 65.2% of chunks as neutral, 19.1% positive and 15.8% negative. The share of negative language differed clearly between banks (20.0%, 11.2% and 9.2%) but stayed steady over time (14.1% to 16.9% per quarter).

The leading themes in negative passages were credit risk (10.0%), treasury and balance-sheet topics (7.1%), cost inflation (4.6%), market volatility (4.1%) and legal risk (3.5%); 36.7% did not match a theme. The pipeline also extracted 358 financial-metric mentions and 345 transcript segments.

Prototype limitations: extracted metrics were not validated against the filings, so they are not used for capital-ratio comparisons. BERTopic was attempted but returned a single topic on this corpus, so themes come from the keyword taxonomy. A forecasting module was drafted but not run; nine quarters per bank is too short for it. Bank names and source files are withheld.

How it was tested

TEST REPORTTEST REPORTG-SIB DISCLOSURESG-SIB DISCLOSURES
Checked
  • Every chunk scored with two independent methods, FinBERT and VADER.
  • Bank, quarter and document metadata kept on every chunk for traceability.
Did not work
  • BERTopic returned a single topic on this corpus, so themes come from the keyword taxonomy.
Limits
  • Extracted financial metrics were not validated against the filings, so they are not used for capital-ratio comparisons.
  • A forecasting module was drafted but not run; nine quarters per bank is too short for it.
  • 36.7% of negative passages did not match a risk theme.

Technology Stack

Python FinBERT VADER GPT-2 BART Hugging Face Transformers Sentence Transformers pandas Plotly Matplotlib
Interested in this work or something similar?