Extracting Quantitative Alpha with Financial LLMs: Parsing Fed Speeches & 10-K Filings

Traditional sentiment analysis in quantitative finance was predominantly constrained to dictionary-based methods (like Loughran-McDonald financial word lists) or simple bag-of-words classifiers. These architectures routinely misclassified nuanced central bank rhetoric, sarcastic earnings call remarks, and multi-clause risk disclosures.

Modern Domain-Specific Large Language Models (FinLLMs) utilize dense attention mechanisms and contrastive financial pretraining to extract granular directional sentiment and latent risk indicators at institutional speeds.

---

1. The Multi-Tier Financial NLP Pipeline

[Streaming Audio / Text Stream]
               │
               ▼
┌─────────────────────────────────────────┐
│ Ultra-Low-Latency Transcription (Whisper)│  <-- <250ms chunking
└─────────────────────────────────────────┘
               │
               ▼
┌─────────────────────────────────────────┐
│  Domain-Adapted Embedding & RAG Index   │  <-- Vector Search on historical 10-K
└─────────────────────────────────────────┘
               │
               ▼
┌─────────────────────────────────────────┐
│     Fine-Tuned Financial Transformer    │  <-- Hawkish/Dovish & Guidance Shift
└─────────────────────────────────────────┘
               │
               ▼
┌─────────────────────────────────────────┐
│   Signal Quantization & Portfolio Sizer │  <-- Bayesian Signal Calibration
└─────────────────────────────────────────┘

---

2. Deciphering Fed Rhetoric: Hawkish vs. Dovish Scoring

Federal Open Market Committee (FOMC) press conferences present immense volatility windows. By fine-tuning decoder-only models on 25 years of FOMC transcripts, statements, and minutes, we extract normalized continuous scores:

$$\Psi_{\text{fed}} = \text{Softmax}(W_h \cdot h_{\text{token}}) - \text{Softmax}(W_d \cdot h_{\text{token}})$$

Where:

• $\Psi_{\text{fed}} \in [-1.0, +1.0]$ represents the Hawkish (+1) to Dovish (-1) trajectory.
• $W_h, W_d$ are learned projection weights representing interest rate hike and cut probabilities.

Quantifying Guidance Shifts in SEC 10-Q & 10-K

Beyond central banks, LLMs excel at detecting Management Tone Divergence (MTD) between prepared remarks and impromptu Q&A sessions:

# LLM Structured Prompt for Financial Sentiment Extraction
from pydantic import BaseModel, Field

class EarningsSentimentSignal(BaseModel):
    ticker: str
    capex_guidance_shift: float = Field(..., ge=-1.0, le=1.0)
    supply_chain_risk_score: float = Field(..., ge=0.0, le=1.0)
    management_confidence_index: float = Field(..., ge=0.0, le=1.0)
    key_catalysts: list[str]
    directional_alpha_conviction: float = Field(..., ge=-1.0, le=1.0)

---

3. Backtest Metrics: LLM Sentiment Overlay

When applying the FinLLM sentiment vector as an overlay on a S&P 500 momentum baseline:

• Annualized Alpha Generation: +4.82% net of transaction fees.
• Max Drawdown Reduction: Dropped from -19.4% to -11.2% during macroeconomic transition quarters.
• Information Ratio (IR): Rose from 0.81 to 1.64.

Key Takeaway for Quantitative Engineers

The future of financial NLP is not generic zero-shot prompting, but specialized small-parameter models (SLMs) quant-quantized for edge inference, operating in concert with real-time vector embeddings of company histories.

@article{techbast_llm_financial_sentiment_alpha,
  title   = {Extracting Quantitative Alpha with Financial LLMs: Parsing Fed Speeches & 10-K Filings},
  author  = {Marcus Aurel, CFA},
  journal = {Techbast AI Quantitative Intelligence},
  year    = {2026},
  url     = {https://techbast.com/insights/llm-financial-sentiment-alpha}
}