Identification of Significant Outliers in Time Series Data
Abstract
This thesis examines the feasibility of using least median of squares (LMS) procedure applied to a reweighted least squares (RLS) autoregression model to identify significant outliers in time series data. The time series were analyzed for data points that were outliers. In order to perform detailed analysis on an outlier. the analyst must be able to determine that an outlier data point is significantly different from normally distributed data. This thesis examines a new method for identifying these outliers. Data from the field were characterized and fit with time series models using an autoregressive reweighted least squares routine (ARRLS) derived from the LMS methodology. Various orders of autoregression were applied to the ARRLS method to determine an appropriate order for the model; resulting fit coefficients were tested for significance. Regression results from data taken at five sites are presented. By using an autoregressive order of one (AR(1)) applied to the ARRL-S, this method significantly improved outlier detection in the time series data over the recursive removal without regression (RRR) method currently in use. In addition to identifying the outliers found by RRR, the AR(1)-RLS method routinely identified four times as many outliers as AFTAC's RRR method. The AR(1)-RLS method is recommended as a complimentary procedure to the RRR method currently used in identifying significant outliers. After sufficient operational experience is gained, AR(1)-RLS may supplant current schemes. Recommendations for improvements to the AR(1)-RLS method are offered.... Outlier, least squares, Autoregression, Least median squared residuals.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 1993
- Accession Number
- ADA262437
Entities
People
- Keri L. Robinson
Organizations
- Air Force Institute of Technology