Missing values in ordered numerical sequences pose a persistent challenge in time-series analysis and sensor data processing. Global imputation methods such as mean and median substitution ignore local structure and signifi cantly distort the distributional properties of the data. We propose the Local Bounded Stochastic Imputer (LBSI), a lightweight framework comprising four variants that impute missing values by sampling within the bounds defined by the nearest non-missing neighbors. The shared bounding constraint en sures every imputed value is locally plausible. The distance-weighted variant(LBSI-D) is proven to be an approximately unbiased estimator for locally lin ear sequences. We evaluate LBSI on two real-world datasets: daily temper ature readings from Delhi (1,462 observations) and hourly carbon monoxide sensor readings from an Italian city (9,357 observations), across four missing rates from 10% to 50%, with 30 Monte Carlo trials per setting. LBSI-Dachieves a Kolmogorov-Smirnov statistic of 0.017 versus 0.284 for mean im putation at 50% missing on the temperature dataset, representing a 16-fold improvement in distributional fidelity.
Roy et al. (Sun,) studied this question.