Methodology

Other Event-Study Types

One design, many outcomes. The classic return event study is only the reference point. Swap the variable you measure, stretch or compress the horizon, and you arrive at a whole family of event-study designs, each with its own estimator and its own valid test. This page lays out that taxonomy.

In short

Every event-study design shares the same four-step skeleton: define an event and an event window, model the counterfactual no-event path over an estimation window, measure the deviation, and test it for significance. What changes from one design to the next is the outcome variable you track, the horizon you track it over, and therefore the estimator and test that keep the inference honest. Below are six market-reaction designs (return, volume, news/sentiment, long-run, intraday, spillover) plus one genuine cousin from a different methodological family, sequence analysis.

The shared event-study skeleton: estimation window, gap, event at t equals zero, event window, and an extended long-run window time estimation window fit the no-event model gap event t = 0 event window extended (long-run) window months to years
The shared skeleton. Every design on this page fits these boxes; each design simply re-scales them and changes what is measured inside the event window.

The taxonomy at a glance

Read the six cards below as a single answer to one question: what is being measured, and over what horizon? Each card states the research question and the outcome it tracks. The full prose for each design follows, and the comparison matrix near the foot of the page lines all six up side by side.

Return (reference)
Did the event change firm value beyond what the market alone would predict? Tracks the abnormal return, signed.

Volume
Did the event change the intensity of trading, regardless of price direction? Tracks abnormal log-volume, unsigned.

News / Sentiment
Does the tone or content of text move prices and volume beyond the hard numbers? Tracks AR or volume conditioned on a text signal.

Long-run
Is there drift or reversal over months to years? Tracks long-horizon abnormal performance (CAR, BHAR, or calendar-time alpha).

Intraday
How fast and how precisely does the price react to a timestamped event? Tracks minute or tick abnormal returns.

Spillover (cross-firm)
When firm A has an event, what happens to connected firms B? Tracks bystander-firm abnormal returns, signed.

Return event study (the reference design)

The return event study is the foundational design and the baseline every other row is "different from". The research question is direct: does the event change firm value? Did prices move more than the market alone would predict? The outcome is the abnormal return, the gap between the realized return and the return you would have expected with no event, cumulated across the window into a CAR for one firm or a CAAR across the sample.

The counterfactual is supplied by a market model fit over an estimation window, commonly 120 to 250 trading days ending before the event:

\[ R_{it} = \alpha_i + \beta_i R_{mt} + \varepsilon_{it}, \qquad AR_{it} = R_{it} - (\hat\alpha_i + \hat\beta_i R_{mt}). \]

Constant-mean-return, market-adjusted, and multifactor (Fama-French) variants all fit the same slot. Inference runs on the CAAR via a cross-sectional or standardized \(t\)-test, with the standardized cross-sectional Boehmer-Musumeci-Poulsen (1991) test to absorb event-induced variance, and the nonparametric Corrado (1989) rank and sign tests as robustness checks. The data requirement is modest: daily adjusted prices or returns for the sample firms, a market index, and the event dates. Seminal references are MacKinlay (1997) and Brown and Warner (1985). The full mechanics are on the introduction to event study methodology page.

Key point

This is the only design that measures a signed change in value at the daily horizon under near-iid, near-normal conditions. It sits in the well-behaved middle of the family. Every other design departs from it on exactly one axis: the variable, the horizon, or the unit of analysis. The site's ARC app implements it.

Volume event study

The volume design asks a different question with the same machinery: did the event change the intensity of trading (information arrival, disagreement, attention), independent of which way the price moved? The outcome is abnormal trading volume, expressed as log-turnover rather than a raw share count. Writing \(V_{it}\) for log-turnover, the fraction of shares outstanding that change hands,

\[ V_{it} = \log\!\left(\frac{\text{shares traded}_{it}}{\text{shares outstanding}_{it}}\right), \qquad AV_{it} = V_{it} - \widehat{E}[V_{it}], \]

with abnormal volume the difference between actual log-turnover and its mean-adjusted or market-model-fitted expectation. The log transform is not cosmetic. Raw volume is strongly right-skewed; Ajinkya and Jain (1989) show that on the log scale it is approximately normal, which is why models are fit on log-turnover and not on raw counts. Testing uses the Campbell and Wasley (1996) parametric and nonparametric (rank) statistics, which are the canonical volume-specific tests because volume violates normality more badly than returns do. Data needed: daily share volume plus shares outstanding to form turnover, plus the event dates. The site's AVC app implements this design. See the sibling volume event study page.

Key point

Volume is one-sided and non-negative: it can only be abnormally high or low, there is no sign of value attached. The log transform and the volume-specific test statistics are mandatory, and the ordinary return \(t\)-test is misspecified on raw volume. Because the signal is unsigned, the volume design is the natural tool for detecting informed or insider trading and run-ups before M&A or news, which is why it recurs in securities litigation and market surveillance.

News / sentiment (content-analysis) event study

This is the design most often missing from a "types" page, yet it is the home turf of a flagship app. The question: does the tone or content of textual information (news, filings, social media) move prices and volume beyond what the hard numbers already imply? The "event" is the publication of text, and the outcome is an abnormal return or abnormal volume conditioned on a text-derived signal: a sentiment or tone score, a topic, or a surprise measure.

The estimator is two-stage. First, quantify the text into a tone or topic score, either by dictionary counts using a finance lexicon or by machine-learning and embedding classifiers. Second, feed that score into a standard return or volume event-study regression or portfolio sort, splitting or weighting the cross-section by the signal. The tests are the same AR, CAAR, or abnormal-volume statistics as in the return and volume designs; predictive regressions of next-day return or volume on tone, in the style of Tetlock (2007), are common. The data are a timestamped text corpus (newswires, 10-K and 8-K filings, earnings-call transcripts, social feeds) aligned to tickers, plus the usual price and volume series. The site's CATA (News Analytics) app is exactly this engine.

Key point

The hard part is measurement of the signal, not the event-study test. The central finance-specific finding is that general-purpose sentiment dictionaries misclassify finance language: Loughran and McDonald (2011) report that roughly three-quarters of the Harvard-IV "negative" words are not negative in 10-K filings. Use a finance lexicon, not an off-the-shelf one. Tetlock (2007) shows that media pessimism predicts both prices and trading volume, tying this design back to both the return and the volume cousins.

Long-run / long-horizon event study

The long-run design asks whether there is drift or reversal after the event over months to years: does the market under- or over-react and then correct slowly? The outcome is long-horizon abnormal performance, typically over 12 to 60 months, and it is measured three competing ways. CAR sums the monthly abnormal returns. BHAR, the buy-and-hold abnormal return, compounds the realized return and subtracts the compounded benchmark or control-firm return. The calendar-time portfolio approach forms a monthly portfolio of all firms currently inside their event window, regresses its excess return on a factor model, and reads the abnormal performance off the intercept:

\[ \text{BHAR}_i = \prod_{t}(1+R_{it}) - \prod_{t}(1+R_{bt}), \qquad R_{pt} - R_{ft} = \alpha + \beta\,(R_{mt}-R_{ft}) + \cdots + \epsilon_t. \]

The benchmark is the methodological hinge. Barber and Lyon (1997) match to control firms by size and book-to-market rather than a raw index, because index benchmarks are misspecified at long horizons. BHAR is tested with the Barber-Lyon skewness-adjusted \(t\) or an empirical bootstrap; the calendar-time alpha is tested with a heteroskedasticity-robust \(t\), which controls the cross-sectional dependence that BHAR ignores. Data: monthly returns over multi-year windows, size and book-to-market characteristics for matching, and factor returns. See the sibling long-run event study page for the full treatment.

Key point

At long horizons the short-window iid and normality assumptions break, and three traps appear: the bad-model problem (small expected-return errors compound with horizon), the skewness of long-horizon returns, and the cross-sectional correlation of overlapping event windows. Fama (1998) argues that most long-run anomalies shrink under reasonable method changes and that the conservative, better-specified calendar-time portfolio is the design to trust; Mitchell and Stafford (2000) show that BHAR overstates significance by ignoring event-firm correlation.

Intraday event study

When the event carries a known timestamp (a press release at 16:01, a Fed decision at 14:00), the question becomes one of speed and precision: how fast and how cleanly does the price react? The outcome is abnormal returns on minute or tick bars within the trading day, and the speed-of-adjustment profile, the cumulative AR by minute. The estimator is a normal or mean-return model calibrated on a short intraday estimation window, for example 60 to 120 minutes, with a 5 to 10 minute gap before the event to avoid leakage. Intraday periodicity in volatility, which is U-shaped over the trading day, must be removed before testing.

Standard parametric tests are misspecified here because minute returns are heavy-tailed, with kurtosis on the order of 5 to 20 against the normal value of 3. The fix is nonparametric or rank-based intraday tests, the heavy-tail rank statistic that the EST intraday tool applies in place of the ordinary \(t\). Data: intraday quote and trade bars (roughly 390 one-minute bars per 6.5-hour US session), with explicit POSIXct timestamps and timezone, aligned across firm, index, and event. Seminal references are Barclay and Litzenberger (1988), who introduced intraday price data to isolate timed announcement effects, and Andersen and Bollerslev (1998) on the intraday volatility footprint of scheduled macro news. See the sibling intraday event studies page.

Key point

Resolution is minutes, not days, so precise event timestamps and timezone alignment are mandatory; timezone misalignment is the single most common practical failure. The payoff is power: Barclay and Litzenberger (1988) report power gains of roughly three to five times for timed events, and Andersen and Bollerslev (1998) document that the bulk of the volatility reaction to scheduled macro announcements is concentrated in the first few minutes after release. The cost is that heavy tails and intraday seasonality force the different test statistics noted above.

Spillover (cross-firm) event study

This is the design the live page has historically called "reverse", and it deserves a precise name. The question: when firm A has an event, what happens to connected firms B (competitors, suppliers, customers, alliance partners, same-industry or same-country peers)? The outcome is the abnormal return, and sometimes volume, of the non-event "bystander" firms, and the sign of that abnormal return is the whole point. A negative sign signals contagion (B falls with A through shared risk or information); a positive sign signals a competitive effect (B rises as it captures the business A loses).

The estimator is the identical market-model AR machinery as the return design, applied to a secondary portfolio of connected firms. Portfolios are formed by industry code, by customer-supplier links from segment data, or by alliance and ownership ties; cross-sectional regressions then explain each bystander's CAR by industry concentration, leverage, or link strength. The tests are the same CAAR \(t\), BMP, and rank statistics, but clustering is severe: bystanders in one industry share the same event date, so portfolio-level or BMP tests are preferred. Data: the event firm's dates plus a map of inter-firm links plus the bystanders' return series. Seminal references are Lang and Stulz (1992) for intra-industry contagion versus competition, and Hertzel, Li, Officer and Rodgers (2008) for supply-chain wealth effects. See the sibling reverse / spillover event study page.

Key point

The unit of analysis is the other firm, not the event firm, and the interesting estimand is the sign and the cross-sectional drivers of the spillover, not merely its existence. Because bystanders cluster on shared event dates, this design fights the same cross-sectional dependence that the long-run design does, and leans on the same portfolio-level remedies.

There is no "volatility event study" type. Variance is not a standalone design. It enters the family in two legitimate but subordinate roles: as a methodological adjustment, the standardized cross-sectional test of Boehmer, Musumeci and Poulsen (1991) that corrects inference for event-induced variance, and as a secondary outcome studied inside the return and intraday designs. Treating "volatility" as a sixth or seventh type confuses an adjustment with a design. Where variance is the object of interest, it is measured as part of a return or intraday study, with the BMP correction applied.

Reverse vs spillover: do not conflate the two. The well-established literature above is the cross-firm / spillover design: a known event at firm A, measured effects on connected firms B. The phrase "reverse event study" is sometimes used in a genuinely different sense, the inverse problem of inferring an unknown event date, or identifying which firms were affected, by working backwards from observed abnormal returns. That inverse problem is what the EDI (Event Date Identifier) app addresses. When you read "reverse event study", check which sense is meant.

Horizon versus power axis, from minutes through days to months and years, with the statistical hazards that grow at each end minutes intraday days return / volume / news months to years long-run power rises, tails fatten well-behaved middle bad-model & dependence grow horizon (log scale, schematic)
Horizon trades off against statistical danger. Shrink to minutes and power climbs but the return tails fatten; stretch to years and the bad-model and cross-sectional-dependence problems take over. The return, volume, and news designs sit in the calm middle.

Comparison matrix

Dimension Return Volume News / Sentiment Long-run Intraday Spillover
Outcome variable Abnormal return (signed) Abnormal log-volume (unsigned) AR / volume conditioned on tone Long-horizon abnormal return Minute / tick abnormal return Bystander-firm AR (signed)
Horizon Days around event Days Days 12 to 60 months Minutes within a day Days around A's event
Sample unit Event firm Event firm Event firm + text Event firm Event firm (timed) Connected firms
Estimator Market / factor model Mean / market model on log-vol Dictionary or ML score + AR model BHAR, CAR, calendar-time \(\alpha\) Short-window normal model Market model on 2nd portfolio
Test statistic \(t\), BMP, Corrado rank Campbell-Wasley parametric + rank As return / volume, split by tone Skew-adjusted \(t\) (BHAR), robust \(t\) on \(\alpha\) Heavy-tail nonparametric rank test CAAR \(t\) / BMP, clustered
Typical data source Daily prices + index Daily volume + shares out. News / filings corpus + prices Monthly returns + size/BM + factors Tick / 1-min bars + timestamps Inter-firm link map + returns

Reading the matrix, three cross-cutting patterns explain why the tests differ. First, signed versus unsigned: the return and spillover designs carry a sign and answer "how much value", while the volume and news designs are unsigned or conditioned cousins that answer "how much information" or "what tone". Second, power versus tails: as the horizon shrinks toward the intraday end, statistical power rises but the return distribution fattens, forcing heavy-tailed tests; the return, volume, and news designs occupy the well-behaved daily middle. Third, dependence: the long-run and spillover designs both fight cross-sectional dependence, from overlapping windows and from clustered events respectively, which is exactly why both lean on portfolio-level remedies such as calendar-time portfolios and the BMP test.

Which EST app fits which design

  • Return → ARC, the Abnormal Return Calculator.
  • Volume → AVC, the Abnormal Volume Calculator.
  • News / Sentiment → CATA, the News Analytics text engine.
  • Long-run → see the long-run event study methodology page.
  • Event-date discovery (the inverse problem) → EDI, the Event Date Identifier.

A note on sequence methods

One genuine cousin sits outside the price-reaction family and is worth flagging precisely because it is a different methodological family, not a market-reaction event study. Sequence analysis asks whether the order and timing of a stream of events, rather than a single shock, carries information: patterns in news-flow, the trajectory of rating changes, or firm and career life-cycle sequences. The method, optimal matching from sociology and bioinformatics (Abbott 1995), encodes each unit as a string of states, computes pairwise edit distances between sequences, and clusters them into types; survival and hazard models are the time-to-event counterpart.

firm A:  calm  calm  shock  recover  calm
firm B:  calm  shock recover calm    calm
         distance = cost of edits aligning the two streams

The outcome here is not a price deviation at all; it is the structure of the event stream itself. We include it as the one row that shows the family extends beyond market-reaction designs, and as the conceptual bridge to news-flow analysis in CATA. It is a social-science sequence method, kept deliberately short so as not to overstate its finance applicability. See the methods for sequence analysis page.

The Event Study Tools site has operated since 2014, building free research tools together with the finance research community. All of the calculators above are free to use.

References

  • MacKinlay, A. C. (1997). Event Studies in Economics and Finance. Journal of Economic Literature 35(1): 13-39. RePEc
  • Brown, S. J. & Warner, J. B. (1985). Using Daily Stock Returns: The Case of Event Studies. Journal of Financial Economics 14(1): 3-31. ScienceDirect
  • Corrado, C. J. (1989). A Nonparametric Test for Abnormal Security-Price Performance in Event Studies. Journal of Financial Economics 23(2): 385-395. ScienceDirect
  • Campbell, C. J. & Wasley, C. E. (1996). Measuring Abnormal Daily Trading Volume for Samples of NYSE/ASE and NASDAQ Securities Using Parametric and Nonparametric Test Statistics. Review of Quantitative Finance and Accounting 6(3): 309-326. Springer
  • Ajinkya, B. B. & Jain, P. C. (1989). The Behavior of Daily Stock Market Trading Volume. Journal of Accounting and Economics 11(4): 331-359. ScienceDirect
  • Tetlock, P. C. (2007). Giving Content to Investor Sentiment: The Role of Media in the Stock Market. Journal of Finance 62(3): 1139-1168. Wiley
  • Loughran, T. & McDonald, B. (2011). When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks. Journal of Finance 66(1): 35-65. Wiley
  • Barber, B. M. & Lyon, J. D. (1997). Detecting Long-Run Abnormal Stock Returns: The Empirical Power and Specification of Test Statistics. Journal of Financial Economics 43(3): 341-372. ScienceDirect
  • Fama, E. F. (1998). Market Efficiency, Long-Term Returns, and Behavioral Finance. Journal of Financial Economics 49(3): 283-306. RePEc
  • Mitchell, M. L. & Stafford, E. (2000). Managerial Decisions and Long-Term Stock Price Performance. Journal of Business 73(3): 287-329. ResearchGate
  • Lang, L. H. P. & Stulz, R. M. (1992). Contagion and Competitive Intra-Industry Effects of Bankruptcy Announcements. Journal of Financial Economics 32(1): 45-60. Semantic Scholar
  • Hertzel, M. G., Li, Z., Officer, M. S. & Rodgers, K. J. (2008). Inter-Firm Linkages and the Wealth Effects of Financial Distress Along the Supply Chain. Journal of Financial Economics 87(2): 374-387. ScienceDirect
  • Barclay, M. J. & Litzenberger, R. H. (1988). Announcement Effects of New Equity Issues and the Use of Intraday Price Data. Journal of Financial Economics 21(1): 71-99. ScienceDirect
  • Andersen, T. G. & Bollerslev, T. (1998). Deutsche Mark-Dollar Volatility: Intraday Activity Patterns, Macroeconomic Announcements, and Longer Run Dependencies. Journal of Finance 53(1): 219-265. PDF
  • Boehmer, E., Musumeci, J. & Poulsen, A. B. (1991). Event-Study Methodology under Conditions of Event-Induced Variance. Journal of Financial Economics 30(2): 253-272. ScienceDirect
  • Abbott, A. (1995). Sequence Analysis: New Methods for Old Ideas. Annual Review of Sociology 21: 93-113. Annual Reviews