Long-Horizon Abnormal Return Inference

A short-window event study — did a stock move abnormally in the three days around an earnings announcement — is statistically well-behaved: returns over a few days are close to normal, and firms rarely overlap in event time. Asking the same question over three or five years, as in "do firms underperform for years after an IPO," breaks nearly every one of those assumptions at once.

Compounding a small daily benchmark-timing error over sixty months multiplies it many times over, so a "buy-and-hold abnormal return" measured over years can differ enormously depending on whether the benchmark is a value-weighted index, a matched control firm, or a characteristic-based portfolio — with published studies on the same event (say, IPO underperformance) reaching contradictory conclusions depending purely on this choice. Overlapping event windows across firms also mean the abnormal returns of different sample firms are not independent draws, since many IPOs or mergers cluster in the same hot market, which inflates apparent statistical significance if treated as independent observations.

Because of these problems, researchers studying long-horizon effects rely heavily on bootstrapped or simulated null distributions (drawing pseudo-portfolios of random firms to build a realistic benchmark for what "normal" long-run variation looks like) rather than trusting a standard parametric t-test, which tends to badly overstate confidence in these settings.