A Comparison Study on Efficiency of Various Time Series Methods for Prediction

• Main Author: 陳森杰
• Other Authors: 高正雄
• Format: Others
• Language: zh-TW
• Published: 2002
• Online Access: http://ndltd.ncl.edu.tw/handle/66531580565968864175

碩士 === 國立中正大學 === 數學研究所 === 90 === The time series model has been widely applied in many aspects, including economics, business, industry, medication, management, and so on. Among those time series that have the relation of function, the dependent series is related not only to the present value but also to the lag value of the independent series. For example, such model can be used to predict sales based on the past and present advertisement expenses and prices, to predict stock prices based on miscellaneous economic indexes, or to predict harvest based on applied amount of fertilizer. It seems that the regression analysis alone cannot completely establish such economic models, for the traditional regression analysis only discusses what input variables influence the output variables and does not take into consideration that the output series can also be possibly influenced by the lag value of the input series. In fact, the transfer function model proposed by Box and Jenkins (1976) can effectively identify the proper model. Accordingly, this thesis work also aims to compare the efficiency of different methods for estimation of impulse response weights used in determining the transfer function models, for cases where there are multiple input variables. In terms of the estimation of weights, there are two categories of methodologies. One is time domain methodology, which consists of the ridge regression method and CCF (cross-correlation function) method. The other is frequency domain methodology, which contains spectral analysis based on Fourier series. The efficiency of estimation is compared between these two methodologies with practical data and simulated data.