Insights

How to Design an Event Study

Alphanume Team · April 25, 2026

Windows, benchmarks, and clean estimation periods — the design choices that determine credibility.

An event study tests whether an identifiable event produces an abnormal return in the stock prices of affected firms. The technique has a defined methodology that has been refined across decades of finance research. Most of the work is in the design — choosing windows, benchmarks, estimation periods, and inclusion criteria — rather than the computation itself.

The seven design choices

An event study requires explicit choices on:

Event definition. What event is being studied, and how is its occurrence identified?
Sample construction. Which firms and which event instances are included?
Event date. What is the specific date associated with each event instance?
Event window. Over what period are abnormal returns measured?
Estimation window. Over what prior period are benchmark model parameters estimated?
Benchmark model. Which model defines expected returns?
Statistical tests. How is significance evaluated?

Event definition and identification

The event must be clearly defined and identifiable from data. Examples:

Earnings announcements (date of release).
Acquisitions (date of announcement, date of closing).
Equity offerings (date of pricing — see 424B5 filing).
Index inclusions (effective date).
Insider trading (filing date of Form 4).

For dilution events, the natural event date is the filing acceptance timestamp of the 424B5 (for shelf takedowns) or the 8-K (for PIPE announcements). Earlier dates (e.g., 8-K announcing the offering before pricing) can be tested separately.

Sample construction

The sample must be:

Complete. All instances of the event in the population, not a curated subset.
Survivorship-bias-free. Include events whose subjects subsequently delisted — see survivorship bias.
Free of confounding events. Exclude events that coincide with other major corporate actions in the event window.
Defined inclusion criteria. Document any filters explicitly.

The choice of period (e.g., 2010–2024 vs 2018–2024) matters. Longer periods produce larger samples but include regime changes; shorter periods are more homogeneous but less powered.

Event date precision

The event date should reflect when the event became known to the market, not when it was scheduled or backdated. For events that release outside trading hours:

Pre-market releases: event date is the same trading day.
After-market releases: event date is the next trading day.
For events that span multiple days: define explicitly whether the event date is the announcement, the pricing, or the closing.

Event window selection

The choice depends on the question:

Information content of the event itself: short windows like (−1, +1) or (0, +1).
Drift and continuation: longer windows like (+1, +20) or (+1, +60).
Pre-event anticipation: windows extending before the event like (−20, 0).

Standard practice is to report results at multiple windows so the time profile is visible.

Estimation window selection

For parametric benchmarks (market model, factor models), the estimation window:

Typical length: 120–250 trading days.
Buffer: 21–30 days between estimation window end and event window start, to avoid leakage from anticipated events.
Skip rules: skip the estimation window if there is insufficient trading data, with consistent handling across the sample.

Benchmark model selection

See how to compute abnormal returns. Brief guidance:

Short windows: market model is standard.
Medium windows: market model or simple factor model.
Long windows: matched-firm or matched-portfolio approaches.

Statistical inference

For cross-sectional inference across the event sample:

T-tests assume normality and constant variance. Often violated.
Standardized CAR tests are more robust to heteroskedasticity.
Bootstrap and permutation tests are robust to distributional assumptions.
For non-normal samples (dilution events typically have heavy tails), non-parametric tests are preferred.

Sensitivity analysis

Good practice includes:

Reporting results at multiple event windows.
Reporting results with and without specific event-type subsamples.
Reporting results across multiple benchmark models.
Reporting median as well as mean.
Examining the cross-sectional distribution, not just the mean.

Common pitfalls

Selection bias. Choosing the sample based on retrospective knowledge.
Confounding events. Other actions in the event window contaminate the result.
Survivorship bias. Excluding delisted names.
Estimation noise leaking into inference. Particularly in factor models with limited estimation data.
P-hacking via window selection. Testing many windows and reporting only the significant ones.

For dilution-event studies specifically

The standard practice for dilution-event studies:

Event date = filing acceptance timestamp of the 424B5 or 8-K.
Event windows: (0, +1), (+1, +20), (+1, +60), (+1, +120).
Estimation window: 250 trading days ending 30 days before event.
Benchmark: market model for short windows; matched-firm (by size and recent return profile) for longer windows.
Subsample on event structure (RDO vs firm-commitment vs ATM vs PIPE).