Signal families

Four independent data families, each evaluated separately before any combination. Definitions, sources, hypotheses, and live evaluation results.

Pre-registered. Signal definitions, evaluation methodology, and failure criteria were committed in PREREGISTRATION.md before any evaluation code ran on real data. Changes are in the amendments log.
Public Attention Wikimedia REST API CC0 / open

Wikipedia Pageviews

Daily article-view counts for each company's primary Wikipedia page. We measure abnormal attention: the z-score of today's views vs. the trailing 60-day window, then cross-sectionally ranked within date and sector.

Hypothesis

H1: Abnormal Wikipedia attention predicts signed and absolute excess returns at 5 and 21 days.

Coverage
S&P 500 + retail basket
Backfill
Historical data available from Wikimedia API (2015–present).
Limitations
Ambiguous company names (e.g. 'Target', 'Block') require manual mapping; precision audited quarterly.
Full evaluation results →
News Sentiment GDELT GKG v2 Open data

GDELT News Tone

Daily article count and average tone from GDELT's Global Knowledge Graph, matched to companies via name/alias dictionaries. Scored with both FinBERT (financial language model) and VADER (lexical baseline) so the two can be compared.

Hypothesis

H2: Abnormal GDELT tone leads returns more in the retail-attention cohort than the S&P 500 core.

Coverage
S&P 500 + retail basket
Backfill
GDELT data available from 2013.
Limitations
Entity matching precision audited; ambiguous names documented in DATA_SOURCES.
Full evaluation results →
Search Interest pytrends (unofficial) Unofficial; cached aggressively

Google Trends

Relative search interest index, normalized with an anchor term to make cross-date comparisons meaningful. Rate-limited; degrades gracefully when blocked. Sampling noise documented.

Hypothesis

H3 (exploratory): Search interest spikes are correlated with Wikipedia attention (corroboration check).

Coverage
S&P 500 + retail basket
Backfill
Historical backfill available with caveats; sampling noise increases for older data.
Limitations
Unofficial API; rate-limited; geographic and language normalization challenges.
Full evaluation results →
Filings Activity SEC EDGAR Public government data

SEC Filings

Form 4 insider transactions, 8-K filing counts, and 13F position changes. EDGAR acceptance timestamps give clean observed_at dates — no look-ahead risk from revision.

Hypothesis

H4: Abnormal insider buy/sell activity predicts 21-day returns.

Coverage
S&P 500
Backfill
EDGAR data available from 1994 for 8-K; 2004 for electronic Form 4.
Limitations
Insiders may transact for non-informational reasons (taxes, diversification, 10b5-1 plans).
Full evaluation results →
No evaluation results yet. The pipeline has been collecting data but sufficient out-of-sample history has not yet accumulated. See the power analysis for the estimated timeline.