Quantitative Analysis of Media Bias and Stock Price Dynamics: The 2020 Shock - Pre-Print Research Paper

Overview

This is a pre-print research paper (arXiv, cs.CE) that I am working on with three co-authors, asking whether the tone of financial news moves stock prices or just reports what the market has already done. We built the full pipeline, from 6.28 million raw news headlines to firm-level sentiment signals to econometric tests, and used the 2020 COVID shock as a natural experiment. The core finding is that the link between news and prices is real but local to individual firms, not a market-wide effect.

The question

Investors, firms and regulators all want to know whether media coverage can move a stock or only mirrors it. Most prior work measures sentiment for the whole market, which cannot tell a genuine market-wide effect from a handful of firms. We asked three questions of 2020: did coverage tone shift, did returns shift, and did the relationship between them change, and in which direction it runs.

Data and approach

We collected 6,284,404 headlines (2015 to 2025) from MediaCloud and narrowed them to 695,731 headlines across 26 US firms, removing thin coverage and near-duplicate wire copy. A fine-tuned DeBERTa-v3 relevance classifier, trained on 750 hand-labelled headlines plus 10,025 synthetic ones, kept only headlines where the firm is the real subject (precision 0.92 at a 0.80 threshold), leaving 90,579. A target-dependent sentiment model then scored each headline's stance toward its firm, and we averaged it into a daily stance measure. We deliberately favoured precision over recall, since a wrong headline distorts the signal more than a missing one does.

Findings

Using fixed-effects panel regressions across 27,627 firm-days, neither media stance nor firm returns shifted after 2020 once each firm was compared with its own history and common daily shocks were removed (p = 0.71 and 0.92). Predictive links did appear, but only firm by firm and after each firm's own structural break: Uber's news led its returns, for example, while for Goldman Sachs the reverse held. Pooled across firms, no market-wide lead-lag channel survived, so the apparent news-to-prices effect looks like a firm-specific story.

Why it matters

For investors, a run of hostile headlines is a signal worth trading on only for specific firms, not the market. For firms and regulators, it argues against treating the press as a lever on the whole market. For analysts building sentiment products, it shows that index-level sentiment hides the firms where predictive signal actually lives.

Methods and tools

• Data engineering: large-scale news collection, deduplication and coverage filtering
• NLP and ML: fine-tuned DeBERTa-v3-base classifier, focal loss, synthetic data augmentation, target-dependent sentiment scoring
• Econometrics: two-way fixed-effects panel regression with robust standard errors, vector autoregressions on firm and pooled panels, Granger causality tests
• Time-series rigour: ADF and KPSS stationarity tests, Bai-Perron structural breaks estimated from the data, Helmert transformation for panel dynamics