请大家帮我看看eviews做二元模型这种结果是什么原因

如果能帮我顺利解决的，按照悬赏给钱哦，麻烦大家了

发完贴为什么看不见图片

看不见图阿

图片文凭先放在附件里了，由于不会贴图，请各位见谅

看不见图啊~！

一看居然是两年前的帖子，解决了吗？

Second Derivative Methods

For binary, ordered, censored, and count models, EViews can estimate the model using Newton-Raphson or quadratic hill-climbing.

Newton-Raphson

Candidate values for the parameters  may be obtained using the method of Newton-Raphson by linearizing the first order conditions  at the current parameter values, :

(32.2)

where  is the gradient vector , and  is the Hessian matrix .

If the function is quadratic, Newton-Raphson will find the maximum in a single iteration. If the function is not quadratic, the success of the algorithm will depend on how well a local quadratic approximation captures the shape of the function.

Quadratic hill-climbing (Goldfeld-Quandt)

This method, which is a straightforward variation on Newton-Raphson, is sometimes attributed to Goldfeld and Quandt. Quadratic hill-climbing modifies the Newton-Raphson algorithm by adding a correction matrix (or ridge factor) to the Hessian. The quadratic hill-climbing updating algorithm is given by:

(32.3)

where  is the identity matrix and  is a positive number that is chosen by the algorithm.

The effect of this modification is to push the parameter estimates in the direction of the gradient vector. The idea is that when we are far from the maximum, the local quadratic approximation to the function may be a poor guide to its overall shape, so we may be better off simply following the gradient. The correction may provide better performance at locations far from the optimum, and allows for computation of the direction vector in cases where the Hessian is near singular.

For models which may be estimated using second derivative methods, EViews uses quadratic hill-climbing as its default method. You may elect to use traditional Newton-Raphson, or the first derivative methods described below, by selecting the desired algorithm in the Options menu.

Note that asymptotic standard errors are always computed from the unmodified Hessian once convergence is achieved.
这段可能有点用。

搜索了一下，看到有人和你同样的问题。
Could someone please help me with this? Thanks
Q1. This is what I got from IGARCH(1,1). Could you please tell me what I should do with this?

Dependent Variable: R_JPY
Method: ML - ARCH (Marquardt) - Student's t distribution
Sample (adjusted): 1/06/1999 1/05/2009
Included observations: 2511 after adjustments
Failure to improve Likelihood after 1 iteration
Unable to evaluate derivatives at current parameter values
Presample variance: backcast (parameter = 0.7)
GARCH = C(2)*RESID(-1)^2 + (1 - C(2))*GARCH(-1)

Variable Coefficient Std. Error z-Statistic Prob.

C -0.004247 NA NA NA

Variance Equation

RESID(-1)^2 1.225047 NA NA NA
GARCH(-1) -0.225047 NA NA NA

T-DIST. DOF 20.00000 NA NA NA

Mean dependent var -0.004247 S.D. dependent var 1.107039

解决方式：
For question 1, it works when I change error distribution as Normal Distribution.

改为正态分布后，这个问题解决了。。

It is difficult to tell without looking into details of your work and data. Specification of the log-likelihood function differs with respect to your choice of error distribution and the solution can get more complicated from the numerical point of view. IGARCH has an additional constraint, which may be too restrictive for the optimization process of the model. On the other hand, it may simply because the errors do not have fat tails.

希望对你有用。
原帖地址，如果感兴趣，可以去看看
http://forums.eviews.com/viewtopic.php?f=4&t=1125