Research note, fourth in the series. Data: SC 13D filings from EDGAR (public domain), prices with delisted issuers included. None of this is investment advice.
The first two autopsies of this series documented two species of the same family of errors: the inclusion look-ahead, which contaminates the list of names that enter the backtest, and the execution look-ahead, which contaminates the price at which the purchase is assumed to have happened. The present note documents the third species, found in an experiment on shareholder activism, and with it the trilogy that motivates the larger work in preparation is closed, since this bias does not live in the list nor in the price, but in something harder to inspect: in who remains alive to be measured. The note is divided into four sections: the activist thesis and the pre-registered design of the experiment; the interim result that looked like a genuine candidate; the survivorship-free test that reduced it to almost nothing; and the residue that survives that test, together with the reason why it does not suffice either.
The activist thesis
An SC 13D filing is the declaration the SEC requires from whoever accumulates more than five percent of an issuer with active intent, that is, with a plan to influence management, and it is worth distinguishing it from the outset from its passive relative, the 13G, which declares the same position without such intent. The academic literature, with the work of Brav and coauthors as the central reference, documents that the target companies of a 13D exhibit a positive drift of weeks after the announcement, a reading that is plausible a priori: the 13D is the public bet of an informed agent who has committed capital and months of campaign.
The experiment started from a census of 61,670 original SC 13D filings submitted between 2006 and 2024, filtered down to events on effectively tradeable issuers. The pre-registration, frozen before observing any return, specified the protections the program adopts as standard: entry at the open of the first trading day after the filing’s acceptance (the lesson of the previous note, incorporated here by design), returns hedged against the market by beta, horizons of 10, 20, 40 and 60 days, inference with clustered errors, year-by-year analysis and correction for multiple comparisons. It should be noted that the null hypothesis of departure, registered in writing, was that the entire effect would live in the overnight gap, as in the rest of the signals derived from regulatory filings that the program has examined.
An interim candidate
The first pass of the experiment was executed on the subset of issuers whose price history was available in the initial ingestion, a subset that, as will be seen, was equivalent to conditioning the sample on the companies that still existed. The results of that pass drew precisely the exception the literature promised: a post-open drift of +0.53% at ten days (clustered t of 3.4), +0.70% at twenty (t 3.1), +0.90% at forty (t 2.6) and +0.87% at sixty (t 2.1). Unlike the cases of the previous notes, here the entry was already the tradeable one and the gap was already discounted, so that the result survived the audits that had killed its predecessors. In the program’s logbook the experiment was recorded, with due caution, as an interim candidate.
However, the test the protocol requires before any verdict remained pending: reconstructing the complete universe with delisted issuers included, a particularly demanding test in this domain because the target companies of activists are, in unusual proportion, small companies that go bankrupt, deteriorate or are acquired.
The test of the delisted
The bias in question is known in the literature as survivorship bias and its mechanics deserve to be stated with care: if the price history only covers the issuers that remain listed, every event whose story ended badly tends to fall out of the sample, not because someone excluded it, but because nobody was left to report it. The resulting sample is not badly constructed in the sense of the previous notes; it is conditioned on the outcome, which is a more silent form of the same sin.
The reconstruction raised the sample to 19,404 events, of which 5,841 correspond to target companies subsequently delisted. With the complete universe, the twenty-day drift fell from +0.70% to +0.10% (t 1.51), below every pre-registered threshold and without surviving the correction for multiple comparisons; at forty days the sign inverted (−0.13%, t 0.23) and at sixty the fall was outright (−0.54%, with the year-by-year analysis unfavorable in all nineteen years evaluated). The medium-term drift, the one the literature celebrates and the interim candidate exhibited, was for the most part an artifact of having measured only the survivors.
It is worth putting on record that the direction of the bias contradicted the intuition registered a priori in the program itself, according to which omitting delisted names would understate the drift, on the understanding that a good part of post-13D delistings are acquisitions paid at a premium. The data point to the contrary: taken together, the target companies that leave the exchange after a 13D have poor subsequent performance (campaigns that fail, issuers that deteriorate until they exit the market), and their absence inflated the average of those that remained. The lesson thus includes a warning about one’s intuitions regarding one’s own biases: they too require measuring.
The ten-day residue
The survivorship test did not kill everything, and rigor obliges documenting what remained. In the complete universe an effect at ten days persists, of +0.31% net of costs (clustered t of 3.51), favorable in all nineteen year-by-year analyses and a survivor of the correction for multiple comparisons. Nevertheless, that residue is concentrated in the first one or two days after the open, and the audit cell the previous note introduced, shifting the entry, dismantles it: entering one day late, the effect is left at +0.09% (t 0.7). A phenomenon of that fragility, whose capture demands executing at the exact open of the correct day on small issuers, belongs to the zone in which real friction (spreads, impact, availability) comfortably consumes what the statistical average promises. It is inferred that the market prices the activist’s campaign at the announcement and its immediate vicinity, and that the holdable drift of weeks, measured where everyone can trade and over everyone who existed, is indistinguishable from zero.
Limitations
The study is an event study with hedged returns per event, not a portfolio backtest, and its net figures assume a fixed cost per trade that in small issuers may fall short, something that reinforces, though not weakens, the negative verdict. The tradeability filter (listing on major exchanges and a minimum of volume) restricts the generalization to the liquid subset of the 13D universe, which is also the only one relevant for practice. The census ends in 2024 by virtue of a change in the SEC’s form nomenclature that affects later ingestion, without consequence for the period analyzed. Finally, the comparison between the surviving subset and the complete universe is presented as a diagnosis of the bias and should not be read as two alternative strategies: the former was never tradeable, which is precisely the point.
Conclusions
The four sections are settled. The activist thesis arrived endorsed by literature and by a pre-registered design that already incorporated the lessons of the previous notes. The interim candidate (+0.70% at twenty days, t 3.1, with tradeable entry) had credentials its predecessors never reached. The test with delisted names reduced it to +0.10% (t 1.51): most of the medium-term drift came from measuring only the survivors, and the sign of that contamination contradicted the program’s own prior intuition. The ten-day residue is real and at the same time untradeable, since it disappears with one day of delay in the entry.
With this piece the trilogy of species is complete: the inclusion look-ahead contaminates who enters the backtest; the execution one, at what price; survivorship bias, who is still there to be counted. The three share the property of producing plausible results, robust to the usual statistical battery, and the three are detected with mechanical audits that require no sophisticated statistics: reconstructing membership, shifting the entry, recovering the dead. The complete taxonomy, with its detection tests, is the object of the working paper in preparation that this trilogy leaves ready.
Series: pre-registered experiments on systematic equity signals. This program publishes methods and negative results; live signals are not published. Data: EDGAR is public domain; prices (EODHD) are used under license and are not redistributed. None of this is investment advice.