臺灣博碩士論文加值系統

時間序列模式構建過程中，模式之選取攸關預測能力之優劣，如何有效、
正確地判定模式，實為一重要之研究課題。近幾十年來許多 學者，諸如
Akaike (1969, 1970, 1973, 1974)、Schwarz (1978)、 Bhansali &
Downham(1977)、Haman & Quinn(1979)、Shibata(1980)等，分別提出各
種不同模式選取準則，然而至今對這些準則仍無較完整之比較與研究。本
研究以CAT, FPE, FPE1, AIC, SBC, BIC1, BIC2, BIC3, BEC, 及 S 等
11 種準則，利用數千組合成資料，配合動差法與最大概似法推估參數，
以其選取正確模式之精確度，探討各準則之性質及適用性。復以 32 組已
知模式之實測資料，探討各準則之適用性。研究結果顯示，整體而言，
以 BIC1準則之選取能力最佳，而CAT準則為最差。

The forecasting accuracy highly depends on the performance of
model selection in time series modeling. Therefore, how to
effectively and accurately identify model is one of the most
important research topics in time series analysis. During
theast few decades, several reseachers, such as
Akaike(1969,1970, 1973,1974)、Schwarz(1978)、Bhansali &
Downham(1977)、Haman & Quinn(1979)、Shibata(1980), proposed
lots of model selection criteria. However, a systemmatical
studies and comparsions on those criteria has not been made.
The aptness and characteristics of those criteria are
investigated by the study of synthetic data based on the
accuracy of selection of ture model in the present study. Both
of the method of moments and maximum likelihood method are used
to estimate the parameters. In the present study, 11 criteria,
CAT , FPE, FPE1, AIC, SBC, BIC1, BIC2, BIC3, BEC, HQ and S are
used to study. Meanwhile, 32 sets of data which model are
well-known are also used to investigate the aptness of those
criteria. The results indicate that BIC1 is the best for the
accuracy, while CAT is the worst.