Signal construction

For each company, we query GDELT GKG for articles mentioning the company name or any registered alias within each trading day. We compute: (1) raw article count, (2) mean GDELT tone field, (3) FinBERT positive probability, (4) VADER compound score. Each is z-scored vs. trailing 60-day window and cross-sectionally ranked.

Limitations & caveats

Entity matching is approximate; precision audited on 200 randomly sampled article-company pairs per source. GDELT includes many marginal articles. Tone in global news may not reflect US equity-relevant sentiment.

Evaluation results

Rank IC (Spearman) with Newey-West standard errors, walk-forward with 21-day embargo. Multiple-testing correction: Benjamini-Hochberg at q=0.05 across all signal × horizon × cohort tests.

No results yet. Data collection is ongoing. Results appear once sufficient out-of-sample history has accumulated. See the power analysis for timeline estimates.

Data coverage

S&P 500 names
—
with ≥1 observation
Retail basket
—
with ≥1 observation
Days collected
—