Historical Analog Imputation: Accuracy, Calibration, and Failure Cases

Historical analogues have the capacity to reconstruct a missing block in time-series data by repeating previously witnessed patterns with similar observed boundaries. Two modifications to this technique are being tested here in GapLens: discarding analog windows that overlap and swapping out a static calibration window with a recent one. A pre-established protocol before scoring measures the performance of fourteen univariate streams, including 1,008 primary masked gaps and 2,139 gaps overall. The exclusion of overlaps makes the equal-series geometric mean absolute error ratio 1.0453 in comparison with the original predictor; it helps in four streams and hurts ten streams. The original predictor is better than linear interpolation in seven streams but is 55.38 times worse in error on one network stream. Using recent calibration shifts, the mean whole gap coverage is decreased from 92.63% to 90.95% on 982 common supported gaps. The apparent reduction in mean interval width under recent calibration is also dominated by one network stream: excluding it in a post-hoc sensitivity analysis changes the recent/static width ratio from 0.3412 to 1.0732.