 |
Dr. Shu-Ing Liu 劉 淑 鶯 |
|
Dr. Shu-Ing Liu 劉 淑 鶯 |
|
時間序列資料之模型建立
時間序列資料在目前是一個相當活躍之研究領域，其應用也很廣泛。資料間彼此相關，因此很多統計推論依賴資料間之獨立假設而得到的，在此不再適用了。而如何建立時間序列資料之模型更是廣泛地被討論。通常一個配適很好的模型，不見得就具有較好之預測能力。 以往之選模準則只箸重於一步預測，以一步預測為目標所決定之模型，在做多期預測時，不具有效性是可以理解的。如果我們有興趣的是同時做多期預測，基於這個需求，如何有效地選取適當的自迴歸模型是最近一項研究成果。依據多期預測之性質，將單變量之自迴歸模型經過適當之轉換後，變成某種形式之多變量迴歸模型。再將幾個較常用之選模準則應用到此『類多變量迴歸模型』。如此即可決定自迴歸模型之秩。數值分析之結果顯示此選模方法確實使得多期預測之結果精確多了。 許多計量經濟之資料具有共整合之現象，雖然成長因素使得各別資料有其一貫之走勢但在長遠均衡狀態下，背後將有某些機制將使得資料間有某些關聯。以線性關聯表示即為現在一般之共整合模型。如何將此模型表現出來，即是一項有趣之研究。其可應用於計量經濟或財務管理上。傳統上，利用正準分析，由特性根來決定模型。我們希望藉由不同研究方向之統計方法來決定模型。目前已有之研究結果有以隨機變數選取方法來決定模型。此方法是仿造迴歸分析中選取變數的方面，每個可能之模型附與機率值，由機率值之大小決定模型之可能性。更進一步，由貝氏分析之觀點，以後驗分配函數來決定較一般化之模型。此方法比傳統之方法更加一般化，傳統之方法是先決定殘差過程之模型，如決定適當之自迴歸模型之秩再決定共整合之秩。而我們的做法則是兩者同時決定，由聯合後驗分配函數來決定。 目前也著手研究資料型態是長記憶之時間序列資料，並考慮有改變點、臨界點及變異數不均勻等之模型之分析，以及多變量非線性模型如
Threshold 模型之檢定等之。
|
|
Model
Building
for Time Series Data Time
series analysis has been an area of considerable activity in recent years.
The in-trinsic nature of a time series is that its obser-vations are
dependent or correlated, and the order of the observations is, therefore
impor-tant. Hence, statistical procedures and tech-niques that rely on
independence assumptions are no longer applicable. Model selection me-thods
for time series data have been discussed by many authors. In general, a
well fitted model may not perform well in prediction. If
we are interested in simultaneous multi-period forecasts, how to select a
suitable au-togressive model is one of our recent research result.
By modifying a univariate auto-regressive model, a suitable
multivariate regression model
was developed in order to create more efficient simultaneous multiperiod
forecasts. Then, some
popularly used variable selection criteria were applied in the
multivariate regression model to develop new selection criteria.
Numerical experiments reveal that the proposed multiperiod
forecasting selection procedure is more efficient than one based on a
one-period forecast. Much econometrics data shows the cointe-gration phenomena. Cointegration means that although many developments can cause permanent changes in the individual elements, there is some long-run equilibrium relation tying the individual components together. Therefore, the representation of such a cointe-grated system is an interesting topic in both economic and financial research fields. Some notes on the method developed are pre-sented in recent technical reports. A new me-thodology was developed to easily build up such a system, by the SSVS method. Moreover, a more general method was developed to simultaneously decide the cointegrated rank and the order of the stationary residual process represented by an autoregressive model. A recent research topic in progress con-centrates on the long-memory data. A wide class of long-memory data is being investigated, include change-point, threshold autoregression and heteroscedastic variance,etc. |
|
Selected
Recent Publications: 1.
Liu,
S. I.(1996)Model
Selection for Multi-period Forecasts.Biometrika,83,861-873. 2. 劉淑鶯,紀穎鴻 (1999), 多重轉折貝氏分析之應用, 中國統計學報, Vol.37, 2, 161-183. 3. Liu, S. I.(2001).Bayesian Model Determination for Binary Time Series Data with Applications. Computational Statistics & Data Analysis,36,461- 473.(#NSC-85-2121-M-008-012) 4. Liu, S. I.(2001).Bayesian Forecasts for Cointegrated Models. to appear in Jourrnal of forecasting. (#NSC- 86-2121-M-008-014)
|