Insights

Abnormal Return: Definition and How to Compute

Alphanume Team · February 26, 2026

Glossary deep-dive with a worked formula.

Abnormal return is the difference between an asset's actual return and the return that would have been expected absent the event being studied. It is the foundational metric of event-driven research — every event-study claim rests on an abnormal-return computation. The mechanics are simple in principle; the choice of expected-return benchmark is where research-quality variation enters.

The basic formula

For a security i on date t:

AR(i, t) = R(i, t) − E[R(i, t)]

Where:

R(i, t) = realized return of security i on date t.
E[R(i, t)] = expected return predicted by the benchmark model.

The worked example

Consider security i with the following data for a single event-window day:

Closing price on day t-1: $50.00
Closing price on day t: $48.00
Realized return: ($48.00 / $50.00) − 1 = -4.00%
S&P 500 return on day t: -1.00%
Security beta (from estimation window): 1.2
Security alpha (from estimation window): 0.0001 (daily)
Predicted return per market model: 0.0001 + 1.2 × (-0.01) = -1.19%
Abnormal return: -4.00% − (-1.19%) = -2.81%

The security underperformed its model-predicted return by 2.81% on the event day. That 2.81% is the day's abnormal return.

Benchmark choices

The expected-return model can be:

Mean-adjusted: Mean return over an estimation window. Simple; doesn't control for market direction.
Market-adjusted: Market return on the event date. Assumes beta of 1.
Market model: Linear regression on market returns. Allows beta and alpha to vary by security.
Multi-factor models: Fama-French 3-factor, 5-factor, Carhart 4-factor.
Matched-firm: Use a control firm or portfolio matched on size, sector, etc.

For most event-study work, the market model is the standard choice. See how to compute abnormal returns for the broader treatment.

The estimation window

For parameter-estimated models (market model, factor models), the estimation window:

Typical: 120-250 trading days ending 21-30 days before the event.
The buffer prevents leakage from pre-event drift contaminating the parameter estimates.

Aggregation

For multi-day analysis, individual ARs aggregate to cumulative abnormal return (CAR):

CAR(i, t1, t2) = Σ AR(i, t) for t in [t1, t2]

For long windows, buy-and-hold abnormal return (BHAR) is often preferred — it compounds rather than sums.

Statistical inference

Single AR computations are starting points; the inference is in the cross-sectional aggregation:

Mean CAR across a sample of events.
Standard deviation across the sample.
T-statistic testing whether mean is significantly nonzero.
Median, distribution analysis for robustness.

Common pitfalls

Look-ahead in benchmark. Using factor loadings computed from post-event data is look-ahead.
Survivorship in sample. Excluding delisted names biases results.
Confounding events. Other events in the window contaminate the signal.
Heteroskedastic returns. Standard t-tests assume constant variance; cross-sectional inference often requires robust standard errors.

Why this metric matters

Abnormal return is the unit of analysis for every claim about event-driven strategies. Post-offering drift is measured in ARs. Lock-up expiration effects are measured in ARs. Survivorship bias and look-ahead bias both manifest as systematic errors in AR computation.

Treating ARs precisely is the foundation of credible event-study research.