New Computational Models In Analysis Of Time Series Data

Abstract

A time series is a collection of observations of a variable taken newlineat regular intervals of time. A forecast, on the other hand, is simply a newlinecalculation of what happens in the future of the variable of interest based newlineon past information under the assumption that the pattern followed in the newlinepast would continue in the future also. newlineThe thesis aims at obtaining forecasting models for the time newlineseries data set using conventional models and computational models. newlineChennai city annual rainfall data for a total of 113 years (1901-2013) has newlinebeen used for the analysis. newlineInitially, individual models are considered and used for newlineforecasting. Later, hybrid models are considered and a comparison newlinebetween individual models and hybrid models are obtained. The newlineindividual statistical models considered are Moving average, newlineExponential smoothing with one parameter and the classical model newlineAutoregressive integrated moving average (ARIMA). Chennai rainfall newlinedata was stationary and therefore differencing of the data series is not newlinerequired whereas the given data was non-normal, hence the data was newlineconverted to make it normal. Many transformations were tried and newlineamong them, natural log transformation considered to be the best. The newlinecriteria for applying ARIMA are stationarity and normality which was newlinemeted out by doing natural log transformation. For doing further analysis newlinethis transformed data was considered for all models including hybrid. In newlinethe case of moving average, for different periods forecast is being carried newlineout and the best is chosen among them with the help of error measures. newlineFor exponential smoothing with a single parameter, the values of the newlinevii newlineparameter are varied and the best among them is chosen based on error newlinemeasures. For ARIMA, tentative models for different values of p, d, q newlineare obtained and among them, the best model for the given data set is newlineidentified with the help of error measures. newlineForecast is also done individually using computational model newlinek-nearest neighbor (KNN) and interpolation technique cubic spline. In newlinethe case of a