AI-powered hedging and evasion detection on live S&P 500 earnings calls. Who's telling the truth?
Every earnings call is theater. CEOs carefully craft their words. But voice doesn't lie the same way text does. Fin45 captures live earnings call audio, transcribes with Whisper large-v3, and runs NLP deception analysis to detect:
Scores range from 0.0 (completely direct) to 1.0 (heavy deception signals). Academic research from Stanford and Wharton shows elevated linguistic deception markers precede negative earnings revisions and stock underperformance by 1-3 quarters.
Deception scoring data will appear here once earnings calls are processed through the NLP pipeline.
The system monitors all S&P 500 earnings calls. Data populates as calls occur and are analyzed.
Unlike services that analyze text transcripts (Seeking Alpha, Bloomberg), Fin45 captures the raw audio stream from earnings call webcasts. Audio is processed through OpenAI's Whisper large-v3 for transcription, then analyzed by a 14B parameter multimodal model (Phi-4) specifically prompted for deception detection.
| Pattern | What It Means | Example |
|---|---|---|
| Hedging increase | Management less certain than they're projecting | "We broadly expect" vs. previous "We will deliver" |
| Q&A evasion | Direct question gets indirect answer | Analyst asks about margins, CEO pivots to revenue growth |
| Passive voice shift | Distancing from negative results | "Revenue was impacted" vs. "We grew revenue" |
| Specificity drop | Previously quantitative, now qualitative | "Strong growth" replacing "$X million increase" |
| Confidence markers | Low confidence despite positive claims | Excessive "I think," "we believe," "it seems" |
Research by Larcker & Zakolyukina (Stanford, 2012) showed CEOs use ~1.5x more hedging words during calls that precede restatements. Hobson et al. (2012) found vocal stress markers during earnings calls predict future SEC investigations. These aren't fringe findings — they're published in the Journal of Accounting Research and The Accounting Review.
A high deception score combined with other bearish signals (insider selling, dark pool distribution, lowered guidance) creates a strong confluence signal for potential short positions or position exits.
Conversely, low deception scores on strong earnings beats (management is being direct AND the numbers are good) confirm bullish conviction.
A 0-1 score measuring hedging language, evasive answers, and tonal inconsistencies in management speech during earnings calls. It uses NLP on actual call audio (not just transcripts) to detect patterns associated with management obscuring information.
No. It means the language patterns match those historically associated with information asymmetry. Executives might hedge for legitimate reasons (regulatory caution, legal advice, genuine uncertainty). The score is most useful when combined with other signals — deception + insider selling is much more concerning than deception alone.
Most tools analyze text transcripts. Fin45 captures live audio and processes it with Whisper + a 14B parameter model specifically prompted for deception markers. Audio contains tonal information that transcripts lose. Additionally, the deception score is one of 11 signal categories feeding into a confluence system — it's not used in isolation.
Earnings calls are processed within hours of completion. During earnings season (Jan-Feb, Apr-May, Jul-Aug, Oct-Nov), new data appears daily as S&P 500 companies report. The table shows the last 90 days of scored calls.