GAUSS的FANPAC

PCGIVE G@RCH

Finmetrics

真的假的，不会这么贵的东东吧。我估计该兄是唬人的

知识无价，更遑论论坛币这种虚拟的咚咚！

ARFIMA、FIGARCH，现在没有一个大家都满意的估计方式，所涉及的技巧，需要的是

Gamma function、 ML estimation、 some theories under time and frequency domain，

换言之，这也是看懂文献的基本要求！

而如果你具备的话， 那还就是回到我说过的，你应该要问的是什麽软件不能拿来估算两类模型，故以 EViews傻瓜软件来当例子！

一堆软件，俯拾皆工具， 因为很好估算嘛！ 预测等等，也不是问题呀！ 不然那些文献是发表假的吗？

反之，你若不具备那些基础，你跑 ARFIMA、FIGARCH 干嘛呢？ 又不懂，到时又引发出一堆涉及计量基础的问题，怎麽估算呀？ 怎麽调整呀？ 怎麽选择呀？ 干嘛，本末倒置的学习呀？

随便玩玩 "回归" 这种大众通俗简单的玩意就好了嘛！ 更何况， ARIMA、ARCH family 还不够玩耍的吗？ 只是 long memory， 多放几个ar terms， "异曲同工之妙"！

就算 能记住较久， 预测较长， 资料也有其限制嘛， 这种预测， 有多少叁考价值呢？

just correct something, Long memory models in fact incorporate both stationary and nonstationary process that is , when d=0 or d=1. Depend on the values you estimate.

Again, long memory is not just for long time dependent. In fact, when you have mixed data like mixed unit root and stationary data, long memory will play an important role in it. For example, monthly inflation data,output growth data, daily stock index data exchange rate data. Sometimes, even for most unit root test, it may misspecify the long memory process as unit root process.

Well, first, do not forget about low power and size distortion among unit roots tests.

Second, the purpose of models with time series analysis, especially in ad hoc method, is mainly on forecast.

If a model or some econometric skill neither incorporate theories well nor forecast reasonably and acceptably, why do we need pay "more" attention on them? Besides, computation cost is always the first in practical.

As you have said, the purpose of the time series modelling is to forecast. So how precise your forecast is depends on how good your model is. Hence developing and finding a better model is most important step during our empirical research.

In fact, there are two groups of research experts. One group is most focus on the theoretical analysis, devoting to the theoretical models. The other group concentrates on the application of the models, that is empirical analyst.

For macroeconomists, they try to model and provide the empirical evidence for monetary policy-making. For financial analyst, they mainly concern with the effect of exogeneous factors, not just time series modelling. In fact, time series models only capture the properties of data itself without any economic meaning.

About the computation cost, sometimes you have no way to avoid it for empirical works. Especially for a new model

you can select

G@RCH 4.0 for GiveWin (Robust)

BaillieBW1996 for WinRats 6.1 (free, but you must install WinRats 6.1)

Fanpac MT 2 for Gauss

Gaussx 7.0

FinMetrics 2.0 for S-plus 7.0

TSMod 4.13 for Ox professional 3.4 (Robust)

Matrixer 5.1 (free, good packages. you can download from http://matrixer.narod.ru/)

You can use Metrixer 5.1 and BaillieBW1996 for WinRats 6.1.

According to hkswen's statement, Long memory effect must be tested by some statistics.

For example, in general, KPSS, GPH, GSE, Lo's R/S, R/V, etc.

In general, Fractionally integrated process can be tested by ARFIMA-FIGARCH model.