Keep Delisted Markets in an Agent Research Denominator
By DX Research Group · · Market data
A four-market fixture shows how a current catalogue can reverse a historical market-selection summary.
Retain removed markets in the historical research population. Filtering an old evaluation through today's surviving catalogue can erase exactly the markets where an agent struggled. We treat delisting as an event in the instrument history, with a documented end to tradability and continuing obligations for any open position.
The Hyperliquid asset-ID documentation ties order identifiers to venue metadata. That current mapping is useful for action routing. An evaluation also needs the archived metadata and market lifecycle relevant to its own period.
An equal-weight result changes sign
Assume four fictional markets, one completed research position per market, equal starting notional, and returns already measured under the same cost convention. This is illustrative arithmetic, with no claim about any venue's actual delistings.
| Market | Position return | In later catalogue? |
|---|---|---|
| A | +6% | Yes |
| B | +2% | Yes |
| C | -8% | No |
| D | -12% | No |
The full population averages (6 + 2 - 8 - 12) / 4 = -3%. Restricting to survivors reports (6 + 2) / 2 = +4%. The seven-percentage-point difference comes entirely from population selection. It demonstrates how a denominator can change the answer, rather than an estimated effect size for real agent trading.
There are several legitimate research populations. A study of current tradable markets can explicitly select current survivors. A study of historical agent performance should preserve historical eligibility. The title, denominator, and exclusions should make that choice visible before a return is interpreted.
Separate removal from unavailable data
A missing price series is an acquisition problem until its cause is established. Preserve a lifecycle reason such as venue removal, renamed market, migrated contract, or unknown. Keep supporting public notices and captured metadata with the reason. Guessing delisting from the absence of a recent candle adds an unsupported event to the dataset.
Where trading stops, record how an open position was valued or settled. Last observed trade, final settlement, and assumed liquidation are different outcome conventions. Carry the convention into the return calculation and show unresolved positions separately. A synthetic terminal price can support sensitivity analysis if it is labelled as an assumption.
Recover the missing population before ranking models
Join decisions to the historical universe, then count instruments with usable entry data, usable exit data, and a complete lifecycle record. Report exclusions by reason and agent. A model whose choices concentrate in removed markets can otherwise lose more observations than another model and appear better merely through selective missingness.
A proposed audit would rerun the summary with all historically eligible markets, followed by a survivor-only sensitivity column. Preserve weights across both calculations so the population change remains interpretable. Our frozen-replay guidance supports binding that population before comparisons. The continuous record shows why bounded populations and explicit execution conventions belong beside results. The fixture here diagnoses a possible selection error; its occurrence in a real dataset requires inspecting the saved market histories.