When Did the Market Know? The Hidden Look-Ahead Problem in Supply-Chain Backtests
September 14, 2026
Altsets
Research by Altsets Research
A relationship can be economically true before it becomes observable, so valid historical testing needs to separate the date the commercial relationship existed from the date a strategy could actually know about it.
Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.
Key findings
- Economic relationship dates and public observation dates can differ, so a historically accurate business reconstruction can still leak future information into a trading backtest.
- Monthly snapshots only solve the problem when they preserve the information state available at that month rather than recomputing historical rows with evidence learned later.
Point-in-time data solves only part of the look-ahead problem. A quant researcher also needs to know when the information became observable.
A commercial relationship can exist economically before it appears in a filing, supplier list, press release, or normalized dataset. If a backtest uses the true economic start date even though the relationship was discovered later, the strategy has been given information the market could not yet have used.
That is a subtle but serious form of leakage.
Economic truth and observable truth can have different dates
Imagine two companies began working together in January but the relationship was not publicly disclosed until April.
A historical database may eventually determine that the commercial relationship existed in January.
For company-history research, January can be the correct economic date.
For a backtest of a public-information strategy, using the relationship in February would be look-ahead bias unless the researcher can establish that the information was observable then.
The correct date depends on the question being tested.
Backfilled relationships are especially dangerous
Relationship datasets improve over time. New filings, archived documents, supplier lists, and entity-resolution work can reveal old commercial connections that were previously unknown.
That creates a temptation to backfill the relationship into the earliest economically plausible period.
Backfilling can be valuable for reconstructing business history. It can be invalid for testing what a trader could have known.
A quant dataset therefore benefits from distinguishing the effective relationship date from the observation or availability date.
Product announcements make the distinction easy to see
A company can announce in 2026 that a current product is being supplied to a customer and also discuss a relationship that began earlier.
The 2026 announcement may strengthen confidence in the historical relationship.
It does not automatically mean a systematic strategy in 2024 was allowed to use the 2026 evidence.
This is why current corroboration should not leak backward into an old test window.
Monthly snapshots are useful only if the snapshot itself is historically honest
A monthly table can still contain look-ahead information if old rows are recomputed with knowledge learned later.
The important question is what each snapshot represents.
Does it represent the best relationship history reconstructed today, or the information state that was available as of that historical month?
Both datasets can be useful, but they answer different research questions.
A trading backtest needs the second one unless the strategy explicitly assumes access to some other contemporaneous source.
Feature timestamps need to propagate through the entire pipeline
The same issue applies after the relationship is ingested.
If a metric is published on May 10, a daily strategy should not use it on May 1. If a monthly snapshot is generated after month-end, the rebalance date should reflect when the snapshot became available.
Lagging the final feature by one period does not automatically fix the problem if the underlying relationship was already backfilled with future knowledge.
The timestamp discipline has to begin at the source.
Leakage can create unrealistically clean lead-lag results
Supply-chain data is naturally attractive for lead-lag research because customers and suppliers report at different times.
That also makes it vulnerable to leakage.
If a relationship is inserted into historical data using information learned after both companies reported, the backtest can appear to identify a clean propagation effect that was not observable in real time.
The stronger the result looks, the more important it is to audit information timing.
The conclusion is to test the information set, not only the relationship history
A historical relationship can be economically correct and still be invalid for a trading backtest.
The quant question is not only "was this relationship true on that date?" It is also "could the strategy have known it on that date?"
The point-in-time backtesting guide covers historical relationship state. The earnings lead-lag guide shows why event ordering and relationship knowledge both need careful treatment.
For relationship definitions and evidence limits, read the Altsets methodology.