Research in progress

Does alternative data
lead stock returns?

Wikipedia pageviews, GDELT news tone, Google Trends, and SEC filings — tracked daily across the S&P 500 and a retail-attention basket. Rigorously tested, honestly reported. Null results included.

Data through Oct 3, 2026 · 503 S&P 500 names · 15 retail-attention names

Live scoreboard

Accumulating history. Evaluation requires sufficient out-of-sample data. The pipeline has been running since Oct 3, 2026. Results will appear here once there is enough data to compute reliable statistics. See the power analysis for the estimated timeline.

Signal families

◈

Wikipedia Pageviews

Abnormal attention spikes in Wikipedia article views for S&P 500 companies. Source: Wikimedia REST API (official, free).
◉

GDELT News Tone

Daily news tone and volume from the GDELT Project's Global Knowledge Graph. Scored with FinBERT and VADER.
◎

Google Trends

Search interest index, normalized relative to an anchor term. Cached with care given sampling noise.
◇

SEC Filings

Insider transactions (Form 4), 8-K counts, and 13F position changes. Clean observed_at from EDGAR acceptance timestamps.
No look-ahead

Every record stores when we observed it and what period it describes. Features for date t may only use data before the t decision time.

Pre-registered

Hypotheses, signal definitions, evaluation metrics, and failure criteria were committed before any evaluation code ran on real data.

Failure shown honestly

Null results appear on their signal pages. A signal that doesn't work says so. The baseline always appears alongside.

Prediction ledger is immutable

Each day's predictions are written and committed before outcomes are known, with model version and data hash. No history rewriting.