Alternative Data for Event-Driven Trading Research

Combine filing activity, public attention, and structured corporate events only after fixing the timestamp and meaning of each input.

Alternative data for event-driven trading is most useful as context around a dated event. Alphanume's SEC Filing Intensity counts issuer filings by day. Wikipedia Views measures public attention relative to a trailing baseline. Structured feeds such as Stock Dilution identify what was disclosed. Joining all three can test whether activity and attention precede an event, but none of the inputs supplies trade direction by itself.

The Filing Intensity documentation defines its simple contract: ticker, company name, date, and filing count. Counts refresh nightly at 11:30 PM Eastern and are available before the next open. Filings submitted after the cutoff appear in the following update. A filing burst measures activity, not whether filings were favorable, material, or dilutive.

Assign a distinct role to each dataset

Use the structured corporate event as day zero. Compute filing activity and attention only from dates before that disclosure. Filing Intensity can show whether the issuer entered an unusually busy reporting period. Wikipedia Views can show whether public attention deviated from its own 30-day baseline. Neither source should be used to label the event type.

Keep every source date in the joined table. A daily attention value should not be aligned to an after-close filing as if both were available before the close. Define whether the feature window ends on the prior calendar day, prior session, or filing timestamp, then apply that rule to every event.

Build a pre-event feature table

Start with dilution events in a completed interval. For each ticker and filing timestamp, calculate filing-count and attention features over a fixed pre-event window. Save the raw daily rows as well as aggregates so unusual values can be traced back to dates. Use cursor pagination for broad Filing Intensity requests by passing both cursor fields.

For each qualifying dilution filing:
1. Set event_time to the filing_timestamp.
2. Pull Filing Intensity for the prior 20 sessions.
3. Pull Wikipedia Views for the same dated window.
4. Calculate filing_count_sum, filing_count_max, views_z_max, and spike-day counts.
5. Keep source dates and missing-day counts.
6. Do not attach post-event returns until the feature table is frozen.

A concrete screen might flag events with an elevated filing count and attention z-score before disclosure. That is a cohort definition, not an edge. The point is to ask whether a pre-event information environment changes the distribution of later outcomes. The event record remains the anchor.

Measure against honest baselines

Raw counts differ across issuers. A company that files frequently needs a different baseline from a quiet issuer, and Wikipedia traffic differs enormously by name. Compare each measure with its own trailing history, then test whether the effect survives size and year splits. Avoid setting thresholds after inspecting return charts.

features["filing_spike"] = features["filing_count_max"] >= 5
features["attention_spike"] = features["views_z_max"] >= 3
features["joint_spike"] = features["filing_spike"] & features["attention_spike"]

audit = features.groupby("joint_spike").agg(
    events=("event_id", "nunique"),
    missing_filing_days=("missing_filing_days", "sum"),
    missing_attention_days=("missing_attention_days", "sum"),
)
print(audit)

Before accepting the thresholds, create placebo cohorts on dates without dilution filings and apply the same feature calculation. This checks whether joint spikes merely identify volatile small companies during busy market periods. Split results by calendar year and market-cap bucket, and keep events that lack one alternative-data source in a separate group. Dropping them can induce coverage bias because thinly followed issuers are exactly where attention records and ticker mappings may be weakest. A feature is useful only if its incremental result survives comparison with the structured event alone and with a predeclared baseline.

Keep the limitations in the result

A filing spike may consist of routine forms. An attention spike may follow a rumor, product event, or broad market move unrelated to the later filing. Coverage gaps can correlate with small or delisted issuers. Multiple filings can represent one financing process, so event de-duplication matters. The Wikipedia Views reference and Stock Dilution guide document the two supporting sources. These features can help segment a population, but causal language requires more than a higher average return in one bucket.

Freeze one cohort and publish its audit

Build one year of dilution events with 20-session pre-event features. Export event keys, filing timestamps, every aggregate, raw-row counts, missingness, and the threshold definitions written before returns. Re-run the pipeline and compare keys. Then attach outcomes and report the unconditional event distribution beside the joint-spike subset. That side-by-side audit is the next action, because it shows whether alternative data adds information or merely removes inconvenient observations.