How Survivorship Bias Can Ruin a Supply-Chain Backtest
September 14, 2026
Altsets
Research by Altsets Research
A historical relationship strategy can look stronger than it was if the universe silently excludes companies and securities that were acquired, delisted, failed, or disappeared before today's dataset was built.
Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.
Key findings
- A current universe excludes securities such as Xilinx that were independently tradable before acquisition, so historical tests need the actual investable universe at each decision date rather than today's survivors.
- Point-in-time relationship research must preserve both economic entities and tradable securities because a security can disappear while the operating relationships continue under a new owner.
A supply-chain backtest can be perfectly coded and still be wrong because the universe itself contains future information.
The easiest version of the mistake is survivorship bias. A researcher downloads today's public companies, retrieves historical prices and relationships for those survivors, and asks whether a supply-chain feature predicted returns ten years ago. The test quietly excludes companies that were acquired, delisted, failed, or otherwise disappeared from the current universe.
That can make a strategy look safer and stronger than it really was.
The historical universe has to contain companies that no longer exist as current securities
Xilinx is a simple example. AMD completed its acquisition of Xilinx in February 2022. A current stock universe no longer contains Xilinx as an independent listed security.
A backtest of semiconductor relationships in 2018 or 2020 should not erase Xilinx merely because today's investor cannot buy the old ticker.
The historical test needs the entities and securities that actually existed at the historical decision date.
This is the same problem fundamental quants face when they build historical universes only from companies that survived into a modern index or vendor database.
Delisting is not the same as economic disappearance
A company can stop trading while its products, customers, suppliers, facilities, and contracts continue under a new owner.
That means a supply-chain dataset needs to distinguish the economic entity from the tradable security.
If the security disappears after an acquisition, the backtest should stop treating the old security as tradable after the transaction. It should not necessarily delete the operating relationships from history.
Otherwise the network history becomes distorted at exactly the corporate events a quant researcher needs to model correctly.
Today's relationship coverage can create a second survivorship problem
There is another version of the bias that is less obvious.
Suppose the dataset has richer relationship coverage for large companies that survived until today and weaker coverage for small companies that disappeared years ago. A strategy built on the well-covered survivors may appear to work because the historical sample quietly favors durable businesses.
The correct test needs to ask what coverage actually existed across the historical universe, including names that later vanished.
Missing relationships should remain missing rather than being interpreted as no dependency.
Point-in-time index membership is only one part of the problem
A researcher can correctly reconstruct historical S&P 500 membership and still introduce survivorship bias through the relationship data.
The security may be point-in-time correct while the supplier and customer network was rebuilt using only companies that exist today.
Both layers need historical integrity.
The test universe, entity mapping, security mapping, relationship state, and data availability all need to line up with the same decision date.
Corporate actions need explicit treatment
Mergers, spin-offs, ticker changes, ADR changes, bankruptcies, and delistings can all alter the mapping between a company and the security used in a strategy.
A robust research pipeline should decide what happens before and after each event.
The goal is not merely to keep prices continuous. The goal is to preserve what the investor could actually trade and what economic relationships belonged to that entity at that time.
This is especially important when a supply-chain signal spans several years.
Survivorship bias can inflate more than returns
The problem can affect concentration measures, factor spreads, hit rates, drawdowns, and network statistics.
A survivor-only sample may appear to have more stable customer relationships, fewer catastrophic supplier events, or stronger long-term relationship persistence simply because the failed companies are missing.
That can lead the researcher to design the wrong feature before a return backtest even begins.
The conclusion is that the historical graph needs a historical universe
A supply-chain backtest should not ask how today's surviving companies would have performed using reconstructed historical features.
It should ask what a researcher could have observed and traded in the actual universe at each historical date.
If disappeared securities and historical entities are missing, the network can inherit survivorship bias before the strategy places its first simulated trade.
The point-in-time backtesting guide explains why current relationship data cannot simply be projected backward. The counterparty-after-delisting guide explains why the operating entity and the security need separate histories.
For relationship definitions and evidence limits, read the Altsets methodology.