Exploring Time Series Randomness

Abstract

Pierpaolo Massoli

Assessing the randomness within time series becomes challenging in the case of large-scale datasets. This novel approach leverages the efficiency of Locality Sensitive Hashing in detecting the repeating patterns over time as well as different time series. By breaking each time series down into pre-defined blocks, the solution set consists of pairs of similar blocks in accordance with the metric the proposed method approximates. As a consequence, the estimation of the aforementioned randomness turns into a pattern recognition problem, insofar as the more patterns are repeated over time, the more predictable the data becomes. Therefore, a simple measurement of the overall randomness of the time series in the input dataset is obtained by counting the identified similar blocks. Following the detection of similar patterns, the mutual information exchanged across the blocks of every detected pair is investigated to validate the results. A case study concerning a selection of different financial market indices is discussed to evaluate the potential of the proposed algorithm.

PDF