When a ten-year return series for an asset class or a portfolio appears on the screen during an investment committee session, the constituents behind that series share one attribute that is written nowhere in the deck: all of them are still trading. Companies present in the set at the start date but which, at some point along the way, lost their listing qualification, entered liquidation, or changed hands under duress and left the exchange do not merely drop out of the series; their final-period return has, in most cases, never been recorded anywhere at all. The curve on the screen measures not the outcome distribution of decisions available over that decade, but the distribution of those that completed the decade — and the distance between those two quantities can be wider than the hurdle rate being argued over in the same room.
The same pattern recurs on non-financial surfaces, usually with less visibility. An institution's approved contractor list does not carry the firms that entered composition proceedings mid-contract or were struck off after losing bonding capacity, which means the historical performance of that list has been pulled upward by the maintenance rule that governs it. A developer's reference file enumerates projects reaching commercial operation; it does not enumerate the ones withdrawn from the interconnection queue, abandoned in permitting, or shelved before FID. In neither case has the record been kept incorrectly — it has been kept so as to represent the present state accurately, and it becomes misleading only when asked a question about probability rather than a question about status.
This pattern carries the name delisting bias — the upward distortion of measured performance produced when units exiting an exchange or a reference set take their returns out of the dataset — and it constitutes a narrower, more traceable, and for exactly that reason more correctable form of survivorship bias. Its mechanism rests on a single structural property: presence in the set is determined by the outcome the analysis is attempting to measure. Membership is not an independent identifying attribute but a dependent variable, so the average of the set is distorted not by sampling error but by the definition of the sample. Increasing sample size does nothing to reduce the distortion, since every additional unit has passed through the same survival filter.
The direction of the distortion appears ambiguous at first, because exits occur for two opposing reasons. A company may leave the set on the upside — through an acquisition at a premium, or by going private at a strong valuation — and failing to record such an exit pulls the measurement downward. On the downside, however, in cases of payment distress, covenant breach, or loss of listing eligibility, terminal value is typically never written at all; it sits in the record system not as a large negative figure but as an empty cell. Because downside exits are both more numerous and heavier in aggregate weight during periods of stress, the net effect runs predictably upward, and it runs furthest upward precisely when the measurement is most needed.
Recognizing that this shortcut is not an error matters for locating the intervention correctly. A register holding only current members is rational not because it is cheap to maintain but because it answers today's question accurately: which instrument can be traded now, which contractor can receive a tender invitation now, which supplier currently satisfies prequalification. The difficulty lies not in the shortcut itself but in the register remaining unchanged when the type of question changes. A record of current state and a record of probability are two distinct documents; in most institutions the second is never opened, so the first is used in its place, and this substitution never surfaces anywhere as a decision.
The first surface on which the cost becomes visible is valuation. When a comparable set is constructed from currently traded peers, its multiple range excludes the peers that went to a discounted sale with a strained balance sheet or disappeared altogether, so the resulting median multiple represents not the outcome distribution of decisions available in that sector on that date but the upper tail of it. The same distortion carries into hurdle rate calibration: a target IRR set against a reference series whose historical performance has drifted upward exceeds what is actually attainable, and in the early years that gap manifests not as performance shortfall but as the rejection of correctly priced transactions — which is to say, it appears in no report at all.
On the project side the distortion bites from a more tangible place. When the contingency reserve of a capital-intensive project is calibrated against the observed cost overrun distribution of completed projects, the right tail of that distribution is structurally truncated, since overruns beyond a certain threshold are not recorded as overruns but as project cancellations or sponsor substitutions and fall outside the overrun series. The same logic repeats in DSCR calibration, in the negotiation of LD caps, and in construction-period insurance pricing: all three are priced against data drawn from projects that reached commercial operation, whereas the risk being priced is precisely the risk of not reaching it.
On the counterparty side the mechanism inverts one step further. The rule governing the approved contractor list mandates removing a firm that falls into payment distress during contract execution; the rule is operationally correct, yet its consequence is to delete from the available sample exactly the event that a team seeking to price contractor insolvency needs to observe. A reference list assembled in a sell-side process carries the same structure, and it does so not because the seller is concealing anything but because a reference list, by construction, enumerates completed work. Buy-side due diligence diverges here: a review that verifies what is in the file and a review that separately requests the list of what never entered the file will produce different prices for the same company.
What neutralizes the distortion is not individual vigilance but record architecture, and it comprises three separable components. The first is an exit register: when a unit leaves the set, its row is not deleted but moved to an archive together with the cause of exit — acquisition, voluntary wind-down, payment distress, compliance loss, cancellation — and its terminal value where one exists, so that an empty cell and a large negative figure become distinguishable. The second is cohort fixing: the comparable set underlying a decision is constructed by reference to membership on the decision date rather than membership today, and every subsequent exit is resolved within that cohort. The third is denominator disclosure: every comparable set carries, immediately beside the multiple or the average, how many units it began with, how many departed and for what reason, and which exclusion rule was applied.
Across BEIREK's project and portfolio lines these three components operate not as a separate document but as a standing annex to the decision process. When an investment or contractor selection decision is opened, the cohort underlying it is locked with a date stamp and read against that same cohort in every subsequent review; every project, supplier, or comparable that leaves the cohort is written into the exit register with its cause and final status rather than removed from the file. When a contingency reserve or schedule buffer is calibrated, the reference set is built not only from completed work but also from work commenced in the same period and never completed, with the difference between the two sets reported as explicitly as the buffer itself.
The second layer concerns rhythm and role separation. Where the party maintaining the exit register is not the party proposing the transaction, an incomplete register becomes a visible gap rather than an omission, and the register is opened at each committee session as a standing agenda item rather than as an annual archival exercise. Scenario material used in stakeholder pre-mortem work draws from the exit register directly rather than from the archive of completed work, since how a project was completed and how a project failed to be completed constitute two distinct bodies of information answering two distinct questions. Establishing that separation improves, before anything else, the ability to discuss which sample a decision was actually taken on.
The genuine performance of a portfolio, a supplier pool, or a sector is read in the file of those who departed rather than in the file of those who remain, and in most institutions that second file is absent not because it was neglected but because keeping it was never decided upon. The question worth asking at a board table is not how good the presented series looks, but how many units the set opened with and how many of them are still in it today.