📞 S&P 500 Earnings Call Sentiment Analysis (VADER + FinBERT)

Turn raw earnings-call transcripts into decision-grade sentiment metrics that quantify how management, analysts, and the market narrative evolve over time.

This repository builds an end-to-end NLP pipeline that produces:

• ✅ Clean speaker-level text blocks
• ✅ Speaker role labelling (Management / Analyst / Operator / Other)
• ✅ Dual-model sentiment scoring
  • VADER (fast, rule-based baseline)
  • FinBERT (finance-domain transformer)
• ✅ Power BI-ready metrics tables
• ✅ A Power BI dashboard: NLP-Dashboard.pbix

---

Why this matters (real business impact)

Earnings calls are not “just text.” They are strategic communication events that influence:

• Investor confidence
• Risk perception
• Market narratives
• Competitive positioning

What sentiment analysis unlocks

Sentiment scoring helps detect changes in:

• Confidence vs caution (tone shifting positive → neutral/negative)
• Uncertainty language (hedging, vague guidance)
• Pressure dynamics (analysts pushing back vs management defending)
• Narrative momentum across quarters and companies

Real-world business advantages

These insights can support:

• Investor Relations (IR): refine messaging; identify where investors are unconvinced
• Equity Research: add consistent sentiment KPIs to qualitative call notes
• Risk / Compliance: flag unusually negative calls for deeper review
• Portfolio strategy: compare narrative trend across companies and time
• Competitive intelligence: benchmark management confidence vs peers

What this proves as an NLP / AI / ML engineering project

This repo demonstrates the core value loop:

Raw text → structured speaker blocks → model scoring → aggregated KPIs → business dashboard

That’s exactly how NLP engineers turn unstructured language into measurable metrics that guide decisions. As an aspiring engineer myself, I ensured to follow this core value loop!

---

Repo Structure

.
├── data/
│   ├── raw/
│   │   └── transcripts_raw.csv
│   └── processed/
│       ├── speaker_blocks_cleaned.csv
│       ├── speaker_blocks_with_vader.csv
│       ├── speaker_blocks_with_finbert.csv          # expected output for merge
│       ├── speaker_blocks_with_sentiment.csv
│       ├── powerbi_call_level_metrics.csv
│       ├── powerbi_role_level_metrics.csv
│       ├── preprocess_checkpoint.txt
│       └── vader_checkpoint.txt
│
├── etl/
│   ├── load_transcripts.py
│   └── preprocess_speaker_blocks.py
│
├── models/
│   ├── sentiment_vader.py
│   └── sentiment_finbert.py
│
├── features/
│   ├── merge_sentimnets.py              # filename typo is intentional (matches repo)
│   └── aggregate_for_powerbi.py
│
├── NLP-Dashboard.pbix
└── requirements.txt

---

Data Source

This project uses the Hugging Face dataset:

• kurry/sp500_earnings_transcripts

etl/load_transcripts.py exports it to:

• data/raw/transcripts_raw.csv

---

Setup

1) Create and activate a virtual environment (recommended)

Windows: python -m venv .venv .venv\Scripts\activate

macOS/Linux: python -m venv .venv source .venv/bin/activate

2) Install dependencies

pip install -r requirements.txt

IMPORTANT: etl/load_transcripts.py uses Hugging Face datasets: from datasets import load_dataset

So install it (and ideally add it to requirements.txt): pip install datasets

3) Install spaCy model

preprocessing uses: spacy.load("en_core_web_sm", disable=["parser", "ner"])

Install: python -m spacy download en_core_web_sm

---

Pipeline (Run Order)

Step 1 — Download transcripts → data/raw/transcripts_raw.csv

python etl/load_transcripts.py

Output:

• data/raw/transcripts_raw.csv

---

Step 2 — Preprocess into clean speaker blocks + role inference

python etl/preprocess_speaker_blocks.py

Outputs:

• data/processed/speaker_blocks_cleaned.csv
• data/processed/preprocess_checkpoint.txt (resume support)

What preprocessing does (technical):

• Reads raw transcripts from data/raw/transcripts_raw.csv
• Parses structured_content safely via:
  • json.loads() first
  • ast.literal_eval() fallback
• Extracts speaker blocks from common keys (segments/blocks/content/dialogue)
• Cleans text:
  • lowercasing
  • removes “forward-looking statements” and “safe harbour” sections
  • strips non-alphabet characters
  • lemmatizes with spaCy
  • removes stopwords
  • keeps alpha tokens with length > 2
• Filters low-signal blocks using:
  • MIN_BLOCK_LEN = 30 words
• Labels speaker role using keyword heuristics:
  • operator / management / analyst / other

Output schema: data/processed/speaker_blocks_cleaned.csv

• symbol
• company_name
• year
• quarter
• date
• speaker
• speaker_role
• clean_text
• block_length

---

Step 3 — VADER sentiment scoring (baseline)

python models/sentiment_vader.py

Outputs:

• data/processed/speaker_blocks_with_vader.csv
• data/processed/vader_checkpoint.txt

Adds:

• sentiment_vader (compound score in [-1, 1])

Notes:

• Uses chunked processing (CHUNK_ROWS = 5000)
• Resume support via vader_checkpoint.txt
• Duplicate-protection: script stops if output exists and checkpoint is 0

---

Step 4 — FinBERT sentiment scoring (finance transformer)

python models/sentiment_finbert.py

Expected output (required for merge step):

• data/processed/speaker_blocks_with_finbert.csv

Required columns (used downstream):

• finbert_sentiment (positive / neutral / negative)
• finbert_confidence (confidence score)

IMPORTANT:

• merge + aggregation scripts assume the file/columns above exist.
• If the current sentiment_finbert.py is not producing them yet, implement/update it so it writes: data/processed/speaker_blocks_with_finbert.csv

---

Step 5 — Merge VADER + FinBERT into one dataset

python features/merge_sentimnets.py

Inputs:

• data/processed/speaker_blocks_with_vader.csv
• data/processed/speaker_blocks_with_finbert.csv

Output:

• data/processed/speaker_blocks_with_sentiment.csv

Merge method:

• strict inner join on: symbol, company_name, year, quarter, date, speaker, speaker_role, clean_text, block_length

---

📊 Power BI Dashboard — Executive Overview

The Python pipeline produces sentiment metrics — but the Power BI report is the “decision layer” that makes those metrics usable in real business workflows.

This project isn’t just NLP in Python — the Power BI layer is where raw text becomes decisions.

The dashboard turns thousands of earnings-call speaker blocks into executive-ready KPIs, allowing business analysts to:

• spot sentiment shifts over time (quarter-to-quarter),
• compare Management vs Analyst tone (credibility vs skepticism),
• separate Prepared Remarks vs Q&A (scripted vs spontaneous),
• and drill into outliers + confidence using purpose-built tooltips.

Business impact (real world):
Sentiment signals can act like an “early-warning system” for guidance risk, investor expectations, PR issues, or competitive pressure — especially when Management’s tone diverges from Analysts’ tone.

---

Exhibit 1 — Executive Overview (Report View)

What the page answers (in 10 seconds)

1. What’s the overall tone right now?
2. Is sentiment improving or deteriorating over time?
3. Is Management more optimistic than Analysts (or vice versa)?
4. Does sentiment change when we move from scripted Prepared Remarks → spontaneous Q&A?
5. What’s the distribution of positive / neutral / negative sentiment?

Filters (left panel)

• Company slicer (single or multi-select)
• Year slicer
• Quarter slicer
• Section slicer: Prepared Remarks vs Q&A
  (this is critical — Q&A is where uncertainty and tension usually shows up)

KPI Banner (top row)

The banner is the “exec summary” for the current filter context:

• Total Calls
• Avg FinBERT
• Avg VADER
• Avg Confidence

Example shown in the screenshot (AAPL):

• Total Calls: 10
• Avg FinBERT: 0.73
• Avg VADER: 0.77
• Avg Confidence: 0.82

---

Exhibit 2 — The Data Backbone (Call-Level vs Role-Level Tables)

These are the two tables that my Python pipeline exports and my Power BI dashboard consumes:

Click to expand (Tables)

Call-Level Metrics Table (one row per earnings call)

Role-Level Metrics Table (one row per call + role + section)

Call-Level table includes:

• volume context (total_blocks, avg_block_len)
• sentiment aggregates (vader_mean, finbert_mean)
• mix metrics (finbert_pos / finbert_neu / finbert_neg)
• reliability signal (finbert_avg_conf)
• divergence signals (vader_gap_mgmt_minus_analyst, finbert_gap_mgmt_minus_analyst)
• traceability ID (CallKey) used in tooltips for “best/worst call” surfacing

Role-Level table includes:

• role splits (Analyst / Management / Operator)
• section splits (Prepared Remarks / Q&A)
• role-level sentiment means + medians (e.g., vader_median)
• role-level confidence

---

Exhibit 3 — Model View (why the tooltips work)

Click to expand (Relational Model)

What’s happening here (technical, but clean)

• Two primary fact tables:
  • powerbi_call_level_metrics_v2
  • powerbi_role_level_metrics_v2
• DimSection supports clean filtering for Prepared Remarks vs Q&A
• A dedicated MeasuresTable centralizes KPI logic and tooltip logic
• Several small tooltip helper tables (TT_*) exist purely to power tooltip layouts:
  • TT Role, TT Sentiment, TT Axis, TT Extremes

This structure allows my tooltips to act as “mini dashboards” without polluting my main model's more minimal overview for a cursory analysis.

---

🔎 The Real Differentiator: Tooltips That Think

I built 3 tooltip pages that activate on hover from the main dashboard. These tooltips don’t just repeat visible charts — they provide diagnostics:

• role divergence
• percentile ranking
• baseline deltas vs company norm
• QoQ momentum
• call extremes (best/worst) with CallKey traceability
• confidence classification (high / medium / low)

---

Tooltip #1 — Role Gap + Gap Percentile (Divergence Diagnostics)

Click to expand (Tooltip #1)

What it shows

• FinBERT Gap and VADER Gap (Management − Analyst)
• Gap Percentile Label (how extreme the divergence is)
• A Role Insights table that breaks down:
  • Sentiment (FinBERT)
  • Sentiment (VADER)
  • Delta (VADER)
  • Delta (FinBERT)
• Confidence badge:
  • ✅ High confidence
  • ⚠️ Medium confidence
  • ❌ Low confidence

Example shown:

• FinBERT Gap = -0.02
• VADER Gap = -0.01
• Gap Percentile Label = 38th%
• Analyst vs Management (FinBERT): 0.80 vs 0.78
• Analyst vs Management (VADER): 0.85 vs 0.84
• Confidence: ✅ High confidence

Why it matters (business)

A divergence between Management and Analysts can signal:

• credibility gaps,
• skepticism in Q&A,
• uncertainty not reflected in scripted remarks,
• narrative management vs fundamentals.

---

Tooltip #2 — Sentiment Mix + “Δ vs Company Baseline”

Click to expand (Tooltip #2)

What it shows

• Distribution cards:
  • +ve Label
  • Neutral Label
  • -ve Label
• Benchmarking cards:
  • vs Company Avg.
  • Δ (change vs baseline)
• A mini chart: Δ vs Company Baseline (how the current context differs from the company’s normal tone)

Example shown:

• +ve Label: 86%
• Neutral Label: 1%
• -ve Label: 13%
• +73% vs Company Avg.
• Baseline deltas show directionality by bucket:
  • Positive: +0.013
  • Negative: -0.004
  • Neutral: -0.009

Why it matters (business)

This prevents the classic analytics mistake:

“This quarter is positive”
instead of
“This quarter is positive relative to this company’s baseline behavior.”

It tells you if the tone is actually unusual or just “business as usual” for that company.

---

Tooltip #3 — QoQ Momentum + Call Extremes (Best/Worst CallKey)

Click to expand (Tooltip #3)

What it shows

• Momentum cards:
  • FinBERT QoQ Δ
  • VADER QoQ Δ
• Scale context:
  • blocks
  • calls
• Sentiment Mix stacked bar (Negative / Neutral / Positive)
• Call Extremes (within Quarter):
  • Best + Worst surfaced using Extreme CallKey
  • includes FinBERT label strength + traceability

Example shown:

• FinBERT QoQ Δ: +0.80
• VADER QoQ Δ: +0.84
• 5,513 blocks
• 175 calls
• Call Extremes:
  • Best: ADBE-20211216-2021Q4 (FinBERT Label: 1.00)
  • Worst: ABT-20220419-2022Q1 (FinBERT Label: 0.06)
• Confidence: ✅ High confidence

Why it matters (business)

This is the “so what?” tooltip:

• what changed (QoQ),
• how strong the change is (distribution),
• and which calls created the movement (extremes).

---

🖱️ Tooltips in Action (Interaction Proof)

These three screenshots show the tooltips activating directly from the Executive Overview page.

Click to expand (Tables)

Tooltip #1 Interaction

Tooltip #2 Interaction

Tooltip #3 Interaction

What these “money shots” prove (to recruiters + stakeholders)

• The dashboard is not static — it’s interactive and context-aware.
• Hovering passes filter context into tooltip pages (company/quarter/section/sentiment bucket).
• The tooltips behave like mini dashboards:
  • Example (AAPL • Q3 • Positive): ❌ Low confidence
    • FinBERT Gap: +0.13
    • VADER Gap: -0.21
    • Gap Percentile: 50th%
    • Role matrix shows cross-model disagreement patterns (FinBERT vs VADER).
  • Example (ABBV): ⚠️ Medium confidence
    • FinBERT QoQ Δ: +0.73
    • VADER QoQ Δ: +0.90
    • 83 blocks, 2 calls
    • Extremes surfaced via CallKey:
      • Best: ABBV-20210203-2020Q4 (0.80)
      • Worst: ABBV-20240202-2023Q4 (0.67)

---

⚡ Key Insights That Can Be Pulled in 60 Seconds (How to actually use this)

Use these as “executive questions” — each one maps directly to a slicer + a visual + a tooltip.

1) “Is sentiment trending up or down — and is it reliable?”

• Where to look: KPI Banner + Earnings Call Sentiment Trend
• What to check:
  • Avg FinBERT / Avg VADER direction over time
  • Avg Confidence (high confidence = signal, low confidence = caution)

2) “Is Management’s optimism aligned with Analyst skepticism?”

• Where to look: Management vs Analyst Sentiment bar chart
• Then hover: Tooltip #1 (Role Gap + Percentile)
• What to check:
  • FinBERT Gap + VADER Gap (Mgmt − Analyst)
  • Gap Percentile Label (is the divergence extreme or normal?)

3) “Does the tone shift when the script ends?”

• Where to look: Prepared Remarks vs Q&A comparison chart
• How: toggle Section slicer (Prepared Remarks vs Q&A)
• What to check:
  • If Q&A sentiment drops while Prepared stays high → potential uncertainty / pressure.

4) “Is this quarter ‘good’… or just normal for this company?”

• Where to look: Hover any relevant context → Tooltip #2 (Δ vs Company Baseline)
• What to check:
  • +ve / Neutral / -ve Label percentages
  • vs Company Avg. uplift
  • Baseline deltas by bucket (Positive/Neutral/Negative)

5) “What changed quarter-over-quarter (momentum), and what drove it?”

• Where to look: Hover the trend or quarter context → Tooltip #3 (QoQ Δ + Extremes)
• What to check:
  • FinBERT QoQ Δ and VADER QoQ Δ
  • blocks + calls (sample size)
  • Call Extremes (Best/Worst CallKey) to identify the exact calls moving the metric

6) “Are the models agreeing or disagreeing?”

• Where to look: Tooltip #1 Role Insights table
• What to check:
  • If FinBERT is strongly positive but VADER is negative (or vice versa), treat as a language-style edge case
  • Use the confidence badge (High/Medium/Low) to decide whether to trust the label

7) “Are we seeing a ‘confidence problem’ (noisy signal)?”

• Where to look: KPI Banner (Avg Confidence) + tooltip confidence badges
• What to check:
  • If confidence is low, prefer directional trends + larger sample contexts
  • Use blocks/calls (Tooltip #3) as a sanity check

8) “Which quarter/company should I investigate first?”

• Where to look: Trend line + Tooltip #3 extremes
• How to decide quickly:
  • pick the quarter with the biggest QoQ swing
  • confirm it has enough blocks/calls
  • grab the Worst CallKey and review the transcript context for that call

---

✅ How to Use the Dashboard (1-minute guide)

1. Pick a Company (or compare multiple).
2. Filter by Year / Quarter.
3. Toggle Prepared Remarks vs Q&A for “scripted vs real”.
4. Hover any key visual to open tooltips:
   • Role Gap diagnostics
   • Baseline deltas vs company norm
   • QoQ change + best/worst calls
5. Use the tooltip insights to answer:
   • “Is sentiment moving?”
   • “Is leadership credibility aligned with analysts?”
   • “Where are the outliers and how confident are we?”

This report is designed with one goal:

Keep the main page clean and executive-friendly — push the deep diagnostics into interactive tooltips.

---

Checkpointing (resume support)

Built for long runs + safe interruption:

• data/processed/preprocess_checkpoint.txt
• data/processed/vader_checkpoint.txt

If interrupted (CTRL+C), rerun the script to resume from the last checkpoint.

---

Troubleshooting

spaCy model missing:

• Error: Can't find model 'en_core_web_sm'
• Fix: python -m spacy download en_core_web_sm

Hugging Face datasets missing:

• Error: No module named 'datasets'
• Fix: pip install datasets

VADER duplicate protection:

• For a clean rebuild, delete:
  • data/processed/speaker_blocks_with_vader.csv
  • data/processed/vader_checkpoint.txt

---

👤 Author

Mitra Boga