1. High-Frequency Text as Event Alpha

Financial news streams (Bloomberg/Reuters) and curated institutional social feeds impact intraday price discovery within milliseconds. Institutional quant desks process millions of headlines per second via streaming transformer pipelines.

2. Event Entity Recognition & Sentiment Scoring

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

class InstitutionalNewsSentimentScorer:
    """
    Real-Time FinBERT News & Headline Sentiment Scorer.
    """
    def __init__(self, model_id: str = "ProsusAI/finbert"):
        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
        self.model = AutoModelForSequenceClassification.from_pretrained(model_id)
        self.model.eval()

    @torch.no_grad()
    def score_headline(self, headline: str) -> float:
        inputs = self.tokenizer(headline, return_tensors="pt", truncation=True, max_length=64)
        logits = self.model(**inputs).logits
        probs = torch.softmax(logits, dim=-1)[0]
        # net polarity = P(pos) - P(neg)
        return float(probs[0] - probs[1])

3. Institutional Summary

  • Exponential Decay Weighting: Headline sentiment alpha exhibits a 15-minute half-life; apply strict exponential time decay weighting $\exp(-\lambda \Delta t)$ when aggregating into trading signals.