What NLP adds
Raw transcripts are useful. Processed transcripts are a dataset. Once you're collecting transcripts systematically (Parts 1–3), you can build longitudinal metrics: Is management getting more cautious? Are they using more hedging language? Did the tone shift between the Q3 and Q4 call? These are signals that don't exist anywhere else.
Four NLP layers
Layer 1 — Confidence scoring
Confident language: specific numbers, definitive statements, "we will," "we expect," concrete timelines
Hedging language: "potentially," "we believe," "subject to," "depending on," "uncertain," "difficult to predict"
Score: (confident_count - hedge_count) / total_sentences × 100
Layer 2 — Management vs analyst tone
Analyze prepared remarks separately from Q&A. Management often sounds more optimistic in prepared remarks. A gap between prepared-remarks confidence and Q&A defensiveness is a signal.
Layer 3 — Quarter-over-quarter language diff
Compare word frequency across calls. Rising frequency of "tariff," "inventory," "macro" signals emerging concerns. Declining mentions of a product line signals deprioritization.
Layer 4 — Keyword alert system
Specific phrases trigger immediate alerts regardless of overall tone: "going concern," "material weakness," "restatement," "withdraw guidance," "SEC investigation."
HEARTBEAT configuration
name: transcript_nlp_pipeline
schedule: "on_new_transcript"
steps:
- load_transcript:
source: "{{ transcript_path }}"
- split_sections:
sections:
- name: prepared_remarks
end_marker: "questions and answers"
- name: qa_section
start_marker: "questions and answers"
- score_confidence:
sections: [prepared_remarks, qa_section]
- extract_keywords:
track:
- "guidance"
- "tariff"
- "inventory"
- "AI"
- "macro"
- "uncertainty"
- "pricing"
- check_red_flags:
phrases:
- "going concern"
- "material weakness"
- "restatement"
- "withdraw guidance"
- "SEC"
- "DOJ"
alert_immediately: true
- compare_to_prior:
quarters: 4
metric: confidence_score
- llm:
prompt: |
Compare this quarter's transcript tone to the prior 4 quarters.
Focus on: confidence trend, keyword frequency shifts, Q&A defensiveness.
Current quarter data: {{ current_metrics }}
Prior quarters: {{ historical_metrics }}
- save:
path: "nlp/{{ company }}_{{ quarter }}_metrics.json"
- notify:
subject: "🧠 NLP Brief: {{ company }} Q{{ quarter }} — Confidence {{ score }}/100"Confidence scorer — Python snippet
import re
CONFIDENT_PHRASES = [
"we will", "we expect", "we are confident", "we delivered",
"record", "exceeded", "outperformed", "strong demand", "accelerating"
]
HEDGE_PHRASES = [
"potentially", "we believe", "subject to", "depending on",
"uncertain", "difficult to predict", "we hope", "if conditions",
"may", "might", "could be", "we cannot guarantee"
]
def score_confidence(text: str) -> dict:
text_lower = text.lower()
sentences = re.split(r'[.!?]', text_lower)
confident = sum(1 for s in sentences if any(p in s for p in CONFIDENT_PHRASES))
hedging = sum(1 for s in sentences if any(p in s for p in HEDGE_PHRASES))
total = len([s for s in sentences if len(s.strip()) > 10])
score = round(((confident - hedging) / total * 100) + 50) if total > 0 else 50
score = max(0, min(100, score))
return {"confidence_score": score, "confident_sentences": confident,
"hedging_sentences": hedging, "total_sentences": total}Red flag detector — Python snippet
RED_FLAGS = [
"going concern", "material weakness", "restatement",
"withdraw guidance", "sec investigation", "doj", "class action",
"liquidity concerns", "covenant violation", "impairment charge"
]
def check_red_flags(text: str) -> list:
text_lower = text.lower()
found = []
for flag in RED_FLAGS:
if flag in text_lower:
# Extract surrounding context
idx = text_lower.find(flag)
context = text[max(0, idx-100):idx+200].strip()
found.append({"flag": flag, "context": context})
return foundQuarter-over-quarter comparison — Python snippet
import json, os, glob
def load_historical_metrics(company: str, quarters: int = 4) -> list:
pattern = f"nlp/{company}_*_metrics.json"
files = sorted(glob.glob(pattern), reverse=True)[:quarters]
history = []
for f in files:
with open(f) as fh:
history.append(json.load(fh))
return history
def trend_analysis(current: dict, history: list) -> dict:
if not history:
return {"trend": "insufficient_data"}
scores = [h.get("confidence_score", 50) for h in history]
avg_prior = sum(scores) / len(scores)
delta = current["confidence_score"] - avg_prior
trend = "improving" if delta > 5 else "declining" if delta < -5 else "stable"
return {"trend": trend, "current_score": current["confidence_score"],
"prior_avg": round(avg_prior, 1), "delta": round(delta, 1)}