Keep missing values separate from zero and invalid data

Four states need four interpretations

A numeric zero can mean an unchanged return or a feature measured at zero. An explicit null means the field has no numeric observation. An absent period means the expected row was not received. A string such as unknown is an invalid type under the report input contract. Collapsing these states makes both quality diagnostics and research conclusions less informative.

JSON Schema distinguishes null from a property that is absent. The report implementation similarly requires the expected row fields and accepts finite numbers or explicit nulls where declared. It rejects numeric strings rather than silently converting them. These choices make input errors visible before a chart presents the result as an observed value.

Work through a small numerical example

Take a synthetic feature sequence of 2, null and 4. The observed-subset mean is three, calculated from two numeric observations. Filling the null with zero changes the mean to two. That change is a modeling assumption about the missing observation, not an innocent way to make a spreadsheet formula run.

At the report’s actual scale, imagine twenty observed rows with two null features. Eighteen values enter the feature distribution and the null share is two divided by twenty, or ten percent. The engine requires at least ten current numeric features for drift. Eighteen passes that computational minimum, but the missingness still requires review and an observed-subset label.

Apply different consequences to different metrics

The drift calculation excludes null current features. It does not construct an extra missing bin or infer the missing feature values. A reviewer must therefore ask whether the observed subset differs systematically from the missing subset. Missingness concentrated in stressed periods can make an apparently stable observed distribution unrepresentative of the intended window.

Performance uses a stricter complete-window rule. One null model or benchmark return withholds the full performance block rather than dropping that month and joining the remaining returns. Feature nulls alone do not invalidate separately complete returns. This separation avoids treating an unrelated feature gap as either permission to invent a return or a reason to erase valid outcome evidence.

Investigate why information is absent

A null can arise from unavailable source coverage, a calculation that was not applicable, a failed transformation or an unresolved correction. The compact fixture does not encode every cause. Keep cause evidence in an investigation record and avoid assigning a reason merely because one explanation would be convenient for the model owner.

For a synthetic incident, compare the raw source field, normalized field and report input for the same date. If the raw value is zero but normalization produced null, the problem is a transformation defect. If the raw source is absent, changing the display cannot resolve the evidence gap. The repair and its review scope differ substantially.

Make exports preserve the distinction

JSON can retain explicit nulls, while a blank CSV cell may be read differently by different programs. A downloadable metrics file should therefore identify withheld calculations and explain null-related reasons. A value of zero must remain zero after export and reimport; it should never disappear through a truthiness check or a missing-value formatting rule.

Test zero, negative values, null, absent rows, invalid strings and an entirely missing return field separately. The report’s original synthetic examples support those software checks. They do not establish the actual missingness mechanism of any investment dataset. A real analysis needs source-specific evidence before imputation, exclusion or a claim that incomplete coverage is harmless.

Keep in the decision record

• Never fill a null with zero without an explicit analytical policy.
• Observed-subset drift is conditional on the values received.
• A return gap withholds the complete performance window.

Source notes

1. JSON Schema — Null and absent properties
2. W3C — Data Quality Vocabulary, Working Group Note, 15 December 2016
3. NIST/SEMATECH e-Handbook — Histogram

Sources inform the stated methods and governance context. All example monitoring observations are synthetic. No regulatory approval or production monitoring certification is implied.

Continue the analysis