PhD Defence – Shilpy Sharma
PhD candidate Shilpy Sharma will defend her thesis "Identification of Overlapping Features in Time Series Data" on October 21, 2015, at 2:00pm in Reynolds 219.
Identification of Overlapping Features in Time Series Data
Time series data can be defined as sample observed sequentially over time and occurs in every field such as finance, stock, bioinformatics and many more. Often time series analysis is the problem of trying to differentiate and extract meaningful statistical patterns that changes in a logical way. It is important to learn the different overlapping patterns in the long term and short term time series data. Conventional methods are based on analysing time series data using only one time series analysis technique.
The principal aim of this Thesis is to examine the overlapping features of time- series data using an improved version of three important time-series analysis techniques such as Locally Weighted Scatterplot Smoothing (LOWESS and the related algorithm LOESS), change point, trend, and control chart patterns. All of the algorithms are tested on real data and data which have been computer-generated.
We first examine smoothing techniques for visualization techniques such as LO(W)ESS. Various techniques are compared for their utility and, for LO(W)ESS, we examine and extend automatic means for smoothing and for outlier. We next analyze the masking effects of change points which can disguise themselves as false trends. By looking at the possible overlap of both phenomena it is possible to obtain a composite picture of the time series when both change point and trend are present. Finally, we compare several techniques for distinguishing macro-effects on the time series. Control charts patterns can be associated with certain assignable causes and recognition of such patterns can accelerate the frequently exhibit variations. Each pattern has special statistical characteristics which differentiate one pattern from another. In our simulations, and in real data, we illustrate that presence of more than one pattern may exist and identification of concurrent pattern is important. Finally, we demonstrate that, for the optimal choice of LO(W)ESS smoothing parameter, attention must be paid to the possibility of a local minimum-maximum error property, and that this focus on the best choice of smoothing parameter may prescribe introducing artificial change points.
Advisors: Charlie Obimbo, David Swayne