Preserve the meaning of gaps

First decide whether the process is modeled on elapsed physical time, local calendar slots, business periods, or event sequence. These grids answer different questions. A missing local hour during a daylight-saving transition is not the same as an unreported ordinary hour. Retain timezone-aware instants and define the business calendar separately. For a regular-grid model, construct the expected slots deliberately so an absent row becomes an explicit missing cell rather than silently compressing elapsed time.

Rates, counters, and interval totals require different aggregation. An instantaneous gauge can be averaged over observed samples, while a cumulative counter may need reset-aware differences. An interval count should be assigned to its covered interval, not treated as a continuous level merely because both are numbers. When holding a gauge constant between reports, state that assumption and cap how long it may remain valid. Otherwise a single stale reading can dominate an hour-long feature without any indication that observation support disappeared.

A missing-value mask records whether a numeric input was actually observed. An age feature records time since its last eligible observation. Together they let the model distinguish a fresh zero from an old value carried forward. A fill value is an interface convenience, not evidence. Fit any learned replacement statistic on eligible training data only, and pass the mask through training and serving with the same convention. Do not let a real zero collide invisibly with the placeholder for missing.

Interpolation uses a structural assumption and, often, both neighboring endpoints. Linear interpolation is causal for a particular decision only after both endpoints are usable by that decision’s cutoff. It can be appropriate for reconstructing an older gap at a later origin while being illegal in a replay of the earlier decision. In pandas 2.3.3, the default linear method treats rows as equally spaced, whereas time interpolation uses time spacing. Neither option automatically enforces the forecasting system’s historical availability boundary.

Treat missingness as potentially informative. A collector may drop samples precisely when a host is overloaded; excluding those intervals can make measured utilization look artificially low. A sales series may be absent because an item was unavailable rather than because demand was zero. Include missingness patterns in regime reports, investigate operational causes, and distinguish a model input gap from an unavailable evaluation label. Each case changes a different part of the learning problem.

Compare candidate policies in a causal replay with deliberate outages of different durations. Measure accuracy, fallback frequency, feature age, and the fraction of targets that remain scoreable. A model that accepts arbitrarily stale filled values can look available while making untrustworthy predictions. Define a staleness threshold that changes behavior explicitly, such as serving a seasonal baseline or returning a degraded-quality flag. Version that policy with the model so an imputation change can be audited like any other deployment change.