Event Window & Estimation Window Selection

Window selection is one of the most important design choices in an event study. The event window determines which days are examined for abnormal returns. The estimation window determines the data used to calibrate the expected return model. Poorly chosen windows can reduce statistical power, introduce bias, or miss the market reaction entirely. This page explains how to select both windows, what gap to use between them, and which window lengths work best for common event types.

Event Study Timeline

An event study divides the time series into three periods. Each plays a distinct role in the analysis.

Day 0 is the event date. Negative days precede the event; positive days follow it. The estimation window comes first, followed by an optional gap, then the event window. These periods must never overlap — overlap would mean the event contaminates the estimation of "normal" behavior.

Event Window

The event window captures the market's reaction to the event. It should be long enough to capture the full reaction (including any information leakage before the announcement and post-event drift) but short enough to avoid confounding from unrelated events.

Trade-offs

Estimation Window

The estimation window provides the historical data used to estimate the expected return model. Its length determines the precision of model parameters (alpha, beta, factor loadings, or GARCH parameters).

The standard choice is 250 trading days (approximately one calendar year). This is the default in most event study software and the recommendation in Campbell, Lo, and MacKinlay (1997).

Gap Between Windows

The gap (also called the guard period) separates the estimation window from the event window. It prevents pre-event information leakage from contaminating the estimation of normal returns.

If the event was anticipated by the market (e.g., through rumors, media speculation, or regulatory filings), stock prices may start moving before the official announcement. Without a gap, these anticipatory movements would be included in the estimation window, biasing the model parameters and reducing the measured abnormal return.

In the EventStudy R package, the gap is controlled by the shift_estimation_window parameter. A value of -10 means the estimation window ends 10 days before the start of the event window.

Recommended Windows by Event Type

Window choices vary by event type. The table below summarizes recommendations from the academic literature and common practice.

How Does Window Length Affect Statistical Power?

Statistical power is the probability of detecting a true abnormal return. Window length affects power through two channels.

Event window length: A longer event window includes more days with zero abnormal returns (noise), diluting the signal. If the event effect is concentrated on days [0, +1], a [-10, +10] window adds 19 days of noise. The cumulative abnormal return (CAR) remains the same, but the variance of the CAR estimate increases, reducing the t-statistic and power.

Estimation window length: A longer estimation window produces more precise estimates of model parameters and the residual variance. This reduces the standard error of abnormal returns, increasing power. The improvement diminishes beyond about 250 days.

For a formal analysis of power as a function of study design parameters, see the Power Analysis & Monte Carlo Simulation page.

How Do I Configure Windows in R?

Windows are specified in the event request table. Each event can have different window parameters.

# Standard M&A event study setup
request <- tibble(
  event_id = 1:3,
  firm_symbol = c("AAPL", "MSFT", "GOOGL"),
  index_symbol = "SP500",
  event_date = c("15.03.2022", "20.04.2022", "10.05.2022"),
  group = "Acquisition",

  # Event window: [-1, +1] (3 days)
  event_window_start = -1,
  event_window_end = 1,

  # Gap: 10 trading days
  shift_estimation_window = -10,

  # Estimation window: 250 trading days
  estimation_window_length = 250
)

task <- EventStudyTask$new(firm_data, index_data, request)
ps <- ParameterSet$new()
task <- run_event_study(task, ps)

# Run with different event windows for robustness
windows <- list(
  narrow = c(-1, 1),
  medium = c(-5, 5),
  wide   = c(-10, 10)
)

results <- map(windows, function(w) {
  req <- request |>
    mutate(event_window_start = w[1], event_window_end = w[2])

  task <- EventStudyTask$new(firm_data, index_data, req)
  ps <- ParameterSet$new()
  task <- run_event_study(task, ps)

  # Extract CAAR
  tidy.EventStudyTask(task, type = "aar")
})

# Compare CAAR across window specifications
bind_rows(results, .id = "window_spec")

Best Practices

• Start with standard windows: Use [-1, +1] for the event window and 250 days for estimation unless you have a specific reason to deviate.
• Include a gap: Always use a gap of at least 5-10 trading days to prevent contamination.
• Run robustness checks: Test your results with at least two alternative event window specifications.
• Report all specifications: If results are sensitive to window choice, report all specifications transparently.
• Check for confounding events: With wider event windows, verify that other firm-specific events do not fall within the window.
• Cite precedent: Reference prior studies in your field that use similar window choices.

Literature

• Campbell, J.Y., Lo, A.W. & MacKinlay, A.C. (1997). The Econometrics of Financial Markets. Chapter 4.
• MacKinlay, A.C. (1997). Event studies in economics and finance. Journal of Economic Literature, 35(1), 13-39.
• Kothari, S.P. & Warner, J.B. (2007). Econometrics of event studies. In Handbook of Corporate Finance, Vol. 1.

What Should I Do Next?

• Data Preparation — format your data for the EventStudy package
• Expected Return Models — choose the right benchmark model
• Choosing the Right Test Statistic — select tests appropriate for your study design
• Power Analysis — simulate power for your specific design