Comparison of Two Data Cleaning Methods as Applied to Volatile Time-Series

Out-of-sample forecasting of historically observed time-series inevitably necessitates the application of a suitable data cleaning method to assist improved accuracy of the obtained results. The existing data cleaning methods though work amply with nonvolatile time series; fail when applied to a volatile time-series. In this paper, the suitability of the k -nearest neighbor approach and sliding window prediction approach is tested on a set of nonvolatile and volatile time-series. The performance comparison is carried out considering the historical record of furniture sales data, PV generation, load power, and ambient temperature data of different time-steps collected from various places in the USA. Further, the effect of parameters allied with both the methods on the preprocessing result is also analyzed. Finally, possible reforms are suggested for the appropriate preprocessing of volatile time-series.