When 48 rows stopped meaning four hours

Forty-eight five-minute intervals make four hours. Forty-eight observed rows do not necessarily do the same. MMG's intraday label builder used a row shift to locate the future price. When market bars were missing, the label could silently refer to a later time than the four-hour forecast named in the interface.

This is the sort of bug that survives plausible model output. The target is still a future price, the rows remain ordered, and a classifier can learn from them. What changes is the question the model is answering. An evaluation can be internally consistent with its labels while those labels violate the product's intended horizon.

Consider a simpler illustrative series observed at 12:00, 12:05, and 12:15. Moving two rows forward from noon lands at 12:15. Looking up a price exactly ten minutes ahead requires a 12:10 observation, which is absent. A row shift returns a value where an exact-time join should report missing data.

The actual continuity audit covered the top 100 items by volume. It inspected 839,687 rows and found 22,830 interior gaps. Of 834,887 labels produced by the 48-row rule, 580,076 landed exactly 14,400 seconds later. That is 69.48% correct. The remaining labels failed the intended elapsed-time contract.

Those counts diagnose the horizon definition. They do not show how much of a previous model score was caused by the error. Answering that would require a matched evaluation under corrected labels. The immediate conclusion was sufficient: a field called four-hour return had to identify a price four real hours away.

The replacement builder joins each observation to the same item's price at the exact future timestamp. Rolling feature windows also restart at gaps. Otherwise, a feature described as covering a recent continuous interval could quietly span missing time even after the target itself was repaired.

This repair deliberately gives up examples that cannot support the requested meaning. Filling a missing future observation with a later trade would restore a row count while reintroducing the horizon problem. Inventing a price would introduce a different assumption. The corrected research pipeline instead preserves the gap as a limit on what can be labeled.

The resulting feature dataset is frozen separately from the operational archive, so the evaluation reads a stable artifact. That makes a metric traceable to particular data and feature rules. It also prevents a later collector update from silently changing the examples behind a recorded result.

The continuity audit and the rebuilt research dataset have different scopes. The audit's 834,887 shifted labels came from its top-100-item sample. The v2 feature dataset contains 802,443 exact-horizon rows across 789 items. Subtracting those counts would not measure how many examples the repair removed.

This is a second, independent reporting trap. Two precise numbers beside the words before and after can imply a controlled comparison that the source never made. The proper denominator for the 69.48% result is the audit's own set of shifted labels. The broader v2 dataset needs its own item count, window, and feature eligibility rules.

Within that v2 dataset, 22.88% of examples met the price-rise target of at least 1%, while 2.19% met the conservative after-tax-positive target. Those rates describe different targets on the corrected research cohort. They should not be borrowed to characterize the earlier daily experiment or the top-100 continuity sample.

A later source audit found a separate problem in the corrected pipeline. It joined future bars before calculating cross-item rank features. Whether an item had a future observation could therefore change another item's present-time feature.

Filtering before ranking is the link between future data and a current input. If Item B lacks a future bar, the old builder drops B from the set of current items used for ranking. In the retained synthetic fixture, Item A's momentum feature percentile changes from 1.0 to 0.0 even though its current history does not change. The repaired builder keeps both current rows and preserves A's 1.0 value.

The September 20 local repair constructs current features before attaching future labels and writes separate training and unlabelled inference artifacts. In the tested synthetic cases, changing future observations no longer changes present inputs or cohort membership. A missing future bar leaves the training target unknown while preserving the current row. Unknown metadata is not filled from later snapshots.

That separation changes the model contract. Historical weights are refused, and no compatible dataset or model has been rebuilt. The public application is unchanged, availability times still involve explicit assumptions, and historical performance has not been recomputed. Fixing the target answered what four hours meant. Fixing the inputs addresses what the model could have known, with separate evidence still needed for how well it predicts.

The transferable check is small: for every retained label, assert that the target timestamp minus the observation timestamp equals the declared horizon, within the same entity. Check feature-window continuity separately. These tests specify meaning rather than merely reproduce the implementation's chosen shift.

The broader lesson is to preserve the unit being measured all the way from data construction to the result paragraph. Rows are observations. Hours are elapsed time. Dataset size belongs to a cohort. Making those distinctions explicit improved the question the experiment could answer before any model architecture changed.