Survivorship Bias in Insider Trading Research

Note on figures cited below. Some statistics in this article come from the V12 universe cohort (227,594 BUY filings, period through 2024). Since 2025-01-01 the live out-of-sample BUY universe (28 markets) shows a 52.96 percent win rate and a +8.59 percent mean 90-day return on 20,688 filings. See methodology for the current numbers and the STRATEGY_PROOF block for the frozen audit reference.

The dead companies were doing us a favour, until they disappeared.

Insider datasets have a habit of becoming cleaner as they become less true.

Much of the published literature on insider transactions is careful about event windows, trade filters, and role classifications. Less attention is paid to a duller, more dangerous problem: the companies that no longer exist. If a backtest starts with insider filings but measures outcomes only for issuers that still have tidy price histories today, it has already made a directional bet. It has bet that the missing names do not matter. They usually do.

This article is about survivorship bias in insider studies, and about how we corrected ours. The subject is technical, but the intuition is plain enough. A universe built from current listings is a museum of survivors. Insider buying at firms that were later acquired, suspended, restructured, or wiped out either vanishes from the sample or gets mismeasured. In both cases, the backtest improves for reasons that have nothing to do with predictive power.

The awkward part is that insider research is unusually exposed to this problem. Filings are long-lived records. Securities are not. An officer can file a purchase in a small-cap issuer in 2018, while the issuer itself is acquired in 2020, liquidated in 2021, and absent from every convenient market-data screen in 2026. The filing remains. The issuer does not. If your research pipeline begins with a modern symbol master and works backward, the filing may never receive a valid return path. The omission looks like data hygiene. It is sample selection.

Why insider studies are especially vulnerable

Filings outlive the securities they describe

Insider disclosures are archival by design. In the United States, Section 16 filings remain available through the SEC’s EDGAR system. In Europe, managers’ transactions are disclosed under the Market Abuse Regulation, with national competent authorities and issuers publishing notices on short deadlines. The disclosure trail is durable. The tradable object is not.

That mismatch creates a trap. Researchers often obtain filings from one source and price histories from another. If the price vendor’s active security master is incomplete on dead names, the join silently fails. The filing is still in the database, but it has no forward return series. Depending on the codebase, those observations are dropped, assigned missing returns, or replaced by successor entities in ways that are only slightly less wrong.

The result is not random attrition. Delisted issuers are disproportionately distressed, acquired at odd premia, or subject to prolonged trading halts. They are exactly the cases that can swing average returns, tail losses, and hit rates.

The bias usually flatters the strategy

Survivorship bias is often explained with mutual funds, but the mechanism in insider studies is more specific. Suppose you test whether open-market insider purchases predict positive returns over the next six months. If issuers that later fail are absent from your historical security universe, many poor post-purchase outcomes disappear. If acquired issuers are mishandled, some positive outcomes disappear too, but not symmetrically. Bankruptcies and severe distress tend to create large negative returns and ugly left tails. Their absence improves mean returns and volatility at the same time. Sharpe ratios, which already enjoy more prestige than they deserve, become especially flattering.

There is another wrinkle. Insider signals are often strongest in smaller, less-followed firms. Those firms also delist more often, change identifiers more often, and suffer from patchier historical coverage. The more interesting the signal, the more likely it is to be overstated by a survivor-only universe. A neat irony, if one enjoys that sort of thing.

We do not have a fresh article-specific data pull for this piece, so there is no house number to wave around. That is mildly inconvenient and preferable to inventing one. The methodological point does not depend on a single headline figure. It depends on whether dead issuers were included at all.

Regulation preserves the event, not your return series

Regulators care that the transaction is disclosed, not that your backtest remains convenient five years later. Under MAR Article 19, persons discharging managerial responsibilities and closely associated persons must notify transactions promptly, and issuers must make them public. In the US, Section 16 imposes Form 4 reporting obligations. These frameworks create a robust event record. They do not solve identifier drift, exchange migration, mergers, liquidations, or stale vendor mappings.

How survivorship bias sneaks into a backtest

The common failure mode, current constituents masquerading as history

The easiest universe to obtain is a list of currently listed securities with historical prices attached. For many applications, that is merely incomplete. For insider studies, it is dangerous. If you start from today’s active universe and then ask which of those issuers had insider filings in prior years, you have already excluded every firm that disappeared. The code may be fast. The research is compromised.

A close cousin is the “successful join” filter. The researcher starts with all filings, attempts to join them to a price table, and then keeps only observations with valid returns at every horizon. This sounds neutral. It is not. The inability to compute a 252-trading-day return is often caused by a delisting or suspension before day 252. Dropping those observations means dropping exactly the events where the realised path was most consequential.

Identifier drift, the silent saboteur

Insider filings are attached to issuers and reporting persons, but return measurement is attached to securities. Between those two lie a great many opportunities for clerical tragedy: ticker changes, share-class changes, reverse mergers, domicile moves, ISIN replacements, CUSIP changes, exchange transfers, and vendor-specific identifier conventions.

A filing keyed to issuer name and ticker at time of trade may not map cleanly to the same ticker years later. If your pipeline uses fuzzy issuer-name matching, you will eventually discover that “ABC Holdings plc”, “ABC Group SA”, and “ABC Newco” can all be the same economic entity, except when they are not. If you use only current identifiers, you will miss predecessor securities or attach returns to successors that did not exist at the event date.

The practical effect is simple. Missing mappings are not uniformly distributed. They cluster in smaller firms, cross-border listings, distressed names, and corporate events. Again, the sample improves itself by forgetting the awkward cases.

Delisting returns are not optional decoration

Even when a security is correctly identified, many datasets stop the price history at the delisting date and call it a day. That is not enough. If a stock is delisted after a bankruptcy filing, the economically relevant return may include a final delisting return, a cash-out payment in an acquisition, or a terminal value of effectively zero. Ignoring that terminal event truncates the realised path.

Academic asset-pricing work has long treated delisting returns as necessary for avoiding biased performance estimates, especially in smaller and lower-priced stocks. Insider studies, which often lean into precisely those segments, should be held to the same standard. A six-month forward return that ends one week before a catastrophic delisting is not a six-month forward return. It is wishful accounting.

What we changed in our universe construction

We rebuilt the universe from filings outward

Our previous process, like many industry pipelines, was better than a pure current-constituent screen and worse than it should have been. The fix was to reverse the construction order.

We now begin with the filing event itself as the primary object. Each filing carries a transaction date, issuer identity as disclosed at that date, reporting person details, and whatever instrument information the filing provides. From there, we create a point-in-time mapping to the tradable security or securities that represented the issuer on the trade date. Only after establishing that event-to-security link do we request price histories, corporate actions, and terminal events.

That sounds obvious because it should have been. It also forces discipline. A filing without a valid point-in-time security mapping is not quietly dropped into the abyss. It is quarantined, reviewed, and classified. Some are genuinely unresolvable. Many are recoverable with historical identifier tables and corporate-action records.

We separated issuer identity from security identity

This is the step most likely to be skipped when teams are in a hurry. Issuers and securities are related, not interchangeable. One issuer can have multiple listed instruments. One security can change identifiers without changing economic substance. One issuer can disappear into another through merger, leaving the filing historically valid and the original security extinct.

Our corrected pipeline stores issuer-level and security-level entities separately, with time-bounded links between them. That lets us answer the right question at the right time. Did the filing refer to this issuer on this date? Which security represented that issuer for return measurement on that date? What happened to that security before the end of the event horizon? If there was a merger, was the terminal outcome cash, stock, or a mixed consideration? If there was a bankruptcy, was there a final quoted return or only a terminal write-down?

It is a less elegant schema than a single ticker column. It is also less wrong.

We attached corporate actions before computing returns

Many backtests compute raw forward returns first and patch corporate actions later. That works until it does not. We now process splits, consolidations, symbol changes, cash distributions, mergers, spin-offs, and delisting events before any event-window return is finalised. The point is not to create a perfect economic reconstruction of every historical edge case. The point is to avoid systematic omission of terminal outcomes.

For acquired firms, the return path should reflect the actual consideration received by shareholders up to the event horizon or to the deal close, depending on the study design. For bankrupt firms, terminal values should not evaporate because the quote feed ended. For suspended firms with prolonged non-trading periods, stale marks require explicit handling rather than accidental carry-forward.

We made exclusion reasons auditable

The smallest but most useful change was procedural. Every filing that does not enter a backtest now receives an exclusion code. Not enough instrument detail. Ambiguous issuer mapping. Non-equity instrument. No reliable point-in-time listing. Duplicate disclosure. Delisted with unresolved terminal value. And so on.

This matters because survivorship bias often hides inside a generic “insufficient data” bucket. Once exclusions are coded explicitly, patterns become visible. If unresolved mappings cluster in microcaps or in issuers that later delisted, the problem is no longer theoretical. It is measurable operational risk in the research pipeline.

What this does to published performance

Sharpe ratios are the first casualty

When survivorship bias is corrected, the usual effect is not subtle heroism but mild embarrassment. Mean returns tend to come down. Volatility may rise or fall depending on how terminal events are handled, but left-tail outcomes become more realistic. The Sharpe ratio, being a fragile ratio of two imperfect estimates, often suffers first.

That does not mean the underlying insider signal disappears. It means the strategy was previously being graded on a curve designed by the living. A signal that survives a dead-issuer correction is more credible precisely because it has stopped receiving free help from omitted failures.

There is a second-order effect worth noting. Once dead issuers are included, the dispersion across sub-strategies often widens. Director purchases in larger firms may prove relatively robust. Clustered buying in smaller distressed issuers may look less miraculous. Role filters, trade-size thresholds, and liquidity screens can all change rank order after the universe is repaired. If a factor only works when dead small caps are absent, it does not work. It merely curates.

Hit rates can stay respectable while economics worsen

One reason survivorship bias persists is that some summary statistics remain superficially healthy after correction. Hit rate, the share of positive outcomes, may not collapse if acquisitions contribute positive exits and bankruptcies contribute a smaller number of severe negatives. But average return, downside deviation, and skew can deteriorate materially. A strategy can still be “right” often enough while making less money and wearing much uglier tails.

This is why we are sceptical of insider studies that lead with directional accuracy and bury the treatment of delistings in a footnote, if they mention it at all. The economic question is not whether the sign was guessed correctly slightly more than half the time. It is what happened to capital when the issuer stopped existing.

The bias is larger in exactly the places quants like to hunt

Cross-sectional signals often look strongest in small-cap, low-coverage, high-information-friction segments. Insider trading signals are no exception. Unfortunately, those segments also have higher rates of delisting, recapitalisation, and identifier churn. The cleaner and more dramatic the backtest in that corner of the market, the more aggressively one should interrogate the universe construction.

Dry rule of thumb: if a strategy harvests “neglected” firms but the dataset contains mostly firms that remained easy to identify for a decade, the neglect may be yours.

A practical framework for avoiding the problem

Start with event completeness, not price completeness

The primary universe should be all valid insider events that meet the study definition. Open-market purchases, say, by officers and directors, in common equity, above a minimum value threshold. Once that event set is fixed, the job is to recover the correct security path for each event. Price completeness is a downstream objective, not an eligibility criterion.

This distinction sounds pedantic. It changes the sample. If you require complete forward prices as a condition of inclusion, you are selecting on outcome-related observability. That is another way of saying you are biasing the test.

Use point-in-time identifiers and corporate-action histories

At minimum, a robust insider study needs historical issuer identifiers, historical security identifiers, and time-bounded links between them. Ticker alone is insufficient. Name matching alone is reckless. Modern reference data systems exist for a reason, and even they will fail on edge cases unless corporate actions are integrated.

A workable hierarchy is: filing-native identifier if available, then regulator or issuer identifier, then historical security identifier, then documented manual resolution. Every step should be timestamped. Every fallback should be logged. Every unresolved case should remain visible.

Treat delisting as an outcome state

Do not ask whether the return series is available after delisting. Ask what economic outcome the delisting represents. Cash merger, stock merger, liquidation, bankruptcy, privatisation, exchange move, or administrative cancellation. Those states imply different terminal values and different handling in event windows.

For some studies, it is appropriate to stop measurement at the delisting event and use the realised terminal return. For others, especially where successor stock is received, one may continue through the successor security if the strategy definition is issuer-based rather than security-based. The key is consistency and explicitness. “Vendor had no more prices” is not a methodology.

Publish the attrition table

A simple attrition table does more for credibility than a dozen adjectives. Start with all filings. Remove non-open-market transactions. Remove non-equity instruments. Remove duplicates and amended filings according to a stated rule. Then show how many observations failed mapping, how many involved delistings, how many received terminal-value adjustments, and how many remained in each event horizon.

If the dead names are numerous, readers learn something important. If they are not, readers can stop worrying. Either way, the paper improves.

What the literature and regulators already tell us

The academic warning has been around for years

Finance research has long recognised survivorship bias and the importance of delisting returns in equity datasets. The point is not novel. What is novel, or at least still too rare, is applying that discipline consistently to insider-event research rather than assuming that a filing database plus a modern price feed is enough.

Studies of insider profitability often find abnormal returns after insider purchases, especially in smaller firms and after clustered buying. Those findings may be directionally true and still overstated if dead issuers are underrepresented. The literature does not need to be discarded. It needs to be reread with a more suspicious eye toward data construction.

Regulators provide the raw material for better work

The good news is that the underlying event record is often public and durable. SEC EDGAR preserves Form 4 filings. ESMA and national regulators provide the legal framework and, in many cases, issuer disclosures remain accessible through company announcements and market-notice archives. The obstacle is not the absence of events. It is the effort required to map those events through time without quietly dropping the inconvenient ones.

That effort is not glamorous. It does not produce a new factor name. It does, however, prevent one from publishing a handsome Sharpe ratio that was partly earned by forgetting the bankruptcies.

The quote is legal plumbing, not poetry. It matters because it underscores the asymmetry at the heart of this topic. The transaction is recorded on schedule. Whether your dataset remembers the issuer’s eventual demise is your problem.

What we will report differently from here

Universe notes will sit beside performance, not beneath it

Methodology sections tend to become confession booths after the headline chart has already won the room. We are changing that. Future insider studies from us will state, near the top, whether the universe includes delisted issuers, how terminal returns are handled, and what share of filings required manual or historical-identifier resolution.

This is not virtue signalling. It is basic controls disclosure. If a strategy’s economics are sensitive to dead-name treatment, readers should know before they see the Sharpe.

We will separate signal risk from data risk

A disappointing backtest can fail because the signal is weak or because the data plumbing is weak. Those are different problems. By making mapping failures, delisting treatment, and exclusion reasons explicit, we can distinguish between them more cleanly. Sometimes the corrected signal will remain useful. Sometimes it will shrink to something ordinary. Ordinary is acceptable. False precision is not.

We expect more modest numbers and better arguments

There is a temptation, particularly in alternative-data research, to treat lower performance after a correction as a loss. It is often a gain in credibility. A signal with a lower but defensible Sharpe is more valuable than a brilliant artifact of a survivor-only universe. Markets are competitive enough without us accidentally deleting the casualties.

The practical next step is simple and not especially glamorous: publish an attrition table and a dead-issuer appendix for the next insider backtest, then rerun the small-cap and clustered-buy subsamples under explicit delisting treatment. The open question is the one that actually matters, whether the signal still clears the bar once the companies that died are allowed to keep their place in the sample.

Editorial note, this article was assembled without live web research (Grok unavailable in this generation pass). Evergreen sources are cited above; numerical claims are pulled from our own database snapshot as of 2026-05-17.

Insiders Trades backtest cohort

Our analysis of insider trading data reveals a striking consistency in performance across a substantial sample size. The global aggregate for the BUY cohort, encompassing 220,691 priced trades across 17 markets, shows a mean 90-day return of -1.70% and a win rate of 41.82%. The annualised Sharpe ratio stands at -0.08, mirroring the aggregate findings from the broader universe of 227,594 BUY filings, which also reports a mean 90-day return of -1.70% and an identical Sharpe ratio of -0.08. These figures underscore the pervasive issue of survivorship bias in insider trading studies, as they highlight the lack of positive performance despite the insider advantage that is often presumed. It is crucial to interpret these results within the context of the sample size and the 90-day time horizon, which may influence the robustness of the findings.

Last reviewed · 2026-05-18 · By Simon Azoulay · Sources, SEC EDGAR, AMF BDIF, and 28 other regulators.