Predictability of Target-Specific Estimation Risk under Partial Temporal Observation

Partial temporal records are often characterized by coverage, yet records with the same coverage can retain different temporal supports and signal values and need not support a target summary equally well. Replaying empirical 24-hour observation masks onto complete accelerometer reference days, we define target-specific realized standardized loss and predict its conditional expectation from a pre-specified ladder spanning coverage, observation structure, retained-signal summaries, and raw partial input. The raw partial-input model improved on the coverage-only reference in all twelve target × condition cells, with all 95% intervals above zero. The largest adjacent gain occurred at the transition to handcrafted partial-observation summaries, whereas the raw partial-input representation did not outperform the summary representation under the frozen architecture. These results show that estimation reliability is not characterized by observation amount alone; temporal support and retained signal evidence also carry predictive information. In a pre-specified descriptive secondary analysis, risk-ranked retention reduced realized loss monotonically across all 36 target × condition × coverage strata.