GARCH模型出现错误：flat log likelihood encountered, cannot find uphill direction

是 否 +2 论坛币

k人 参与回答

经管之家送您一份

应届毕业生专属福利!

求职就业群

赵安豆老师微信：zhaoandou666

经管之家联合CDA

送您一个全额奖学金名额~ !

立即领取

感谢您参与论坛问题回答

经管之家送您两个论坛币！

+2 论坛币

我的数据是上市公司十年4921条研发投入数据，时间有gap，但能保证连续7年及以上有数据，想要做的是对每家公司进行分组回归，构建garch模型，求出研发投入变动趋势，出现错误：flat log likelihood encountered, cannot find uphill direction，我猜测可能原因是数据不够，但是已经发表的文章上用类似的数据也做出了结果。所以我想请问各位大佬有什么解决办法吗？这是我的代码：

1. use c:\users\administrator\desktop\7year.dta,clear
   
2. tsset company year
   
3. egen g=group(company)
   
4. sum g
   
5. local n=r(max)
   
6. gen f=.
   
7. forv i=1/`n'{
   
8. arch rd L.rd if g==`i',arch(1) garch(1)
   
9. predict e if e(sample),r
   
10. replace f=e if e(sample)
    
11. drop e
    
12. }
    
13. drop g

复制代码

部分结果

1. Iteration 344: log likelihood = -100.04287  
   
2. (switching optimization to BHHH)
   
3. Iteration 345: log likelihood =  -100.0428  (not concave)
   
4. Iteration 346: log likelihood = -100.04267  (not concave)
   
5. Iteration 347: log likelihood = -100.04267  (not concave)
   
6. Iteration 348: log likelihood = -100.04267  (not concave)
   
7. Iteration 349: log likelihood = -100.04267  (not concave)
   
8. (switching optimization to BFGS)
   
9. BFGS stepping has contracted, resetting BFGS Hessian (87)
   
10. Iteration 350: log likelihood = -100.04267  
    
11. Iteration 351: log likelihood = -100.04267  (backed up)
    
12. Iteration 352: log likelihood = -100.04266  (backed up)
    
13. BFGS stepping has contracted, resetting BFGS Hessian (88)
    
14. Iteration 353: log likelihood =  -100.0425  
    
15. Iteration 354: log likelihood = -100.04223  (backed up)
    
16. Iteration 355: log likelihood = -100.04223  (backed up)
    
17. BFGS stepping has contracted, resetting BFGS Hessian (89)
    
18. Iteration 356: log likelihood = -100.04142  
    
19. Iteration 357: log likelihood = -100.04134  (backed up)
    
20. Iteration 358: log likelihood = -100.04133  (backed up)
    
21. BFGS stepping has contracted, resetting BFGS Hessian (90)
    
22. Iteration 359: log likelihood = -100.04061  
    
23. (switching optimization to BHHH)
    
24. Iteration 360: log likelihood = -100.04058  (not concave)
    
25. Iteration 361: log likelihood = -100.04051  (not concave)
    
26. Iteration 362: log likelihood = -100.04051  (not concave)
    
27. Iteration 363: log likelihood = -100.04051  (not concave)
    
28. Iteration 364: log likelihood = -100.04051  (not concave)
    
29. (switching optimization to BFGS)
    
30. BFGS stepping has contracted, resetting BFGS Hessian (91)
    
31. Iteration 365: log likelihood = -100.04051  
    
32. Iteration 366: log likelihood = -100.04051  (backed up)
    
33. Iteration 367: log likelihood = -100.04051  (backed up)
    
34. BFGS stepping has contracted, resetting BFGS Hessian (92)
    
35. Iteration 368: log likelihood =  -100.0403  
    
36. Iteration 369: log likelihood = -100.04014  (backed up)
    
37. Iteration 370: log likelihood = -100.04014  (backed up)
    
38. BFGS stepping has contracted, resetting BFGS Hessian (93)
    
39. Iteration 371: log likelihood = -100.03937  
    
40. Iteration 372: log likelihood = -100.03935  (backed up)
    
41. Iteration 373: log likelihood = -100.03934  (backed up)
    
42. BFGS stepping has contracted, resetting BFGS Hessian (94)
    
43. Iteration 374: log likelihood =  -100.0388  
    
44. (switching optimization to BHHH)
    
45. Iteration 375: log likelihood = -100.03878  (not concave)
    
46. Iteration 376: log likelihood = -100.03869  (not concave)
    
47. Iteration 377: log likelihood = -100.03869  (not concave)
    
48. Iteration 378: log likelihood = -100.03869  (not concave)
    
49. Iteration 379: log likelihood = -100.03869  (not concave)
    
50. (switching optimization to BFGS)
    
51. BFGS stepping has contracted, resetting BFGS Hessian (95)
    
52. Iteration 380: log likelihood = -100.03869  
    
53. Iteration 381: log likelihood = -100.03869  (backed up)
    
54. Iteration 382: log likelihood = -100.03869  (backed up)
    
55. Iteration 383: log likelihood = -100.03864  
    
56. BFGS stepping has contracted, resetting BFGS Hessian (96)
    
57. Iteration 384: log likelihood = -100.03859  
    
58. Iteration 385: log likelihood = -100.03859  (backed up)
    
59. Iteration 386: log likelihood = -100.03857  (backed up)
    
60. BFGS stepping has contracted, resetting BFGS Hessian (97)
    
61. Iteration 387: log likelihood = -100.03836  
    
62. Iteration 388: log likelihood = -100.03826  (backed up)
    
63. Iteration 389: log likelihood = -100.03826  (backed up)
    
64. (switching optimization to BHHH)
    
65. Iteration 390: log likelihood = -100.03755  (not concave)
    
66. Iteration 391: log likelihood = -100.03748  (not concave)
    
67. Iteration 392: log likelihood = -100.03748  (not concave)
    
68. Iteration 393: log likelihood = -100.03748  (not concave)
    
69. Iteration 394: log likelihood = -100.03748  (not concave)
    
70. (switching optimization to BFGS)
    
71. BFGS stepping has contracted, resetting BFGS Hessian (98)
    
72. Iteration 395: log likelihood = -100.03748  
    
73. Iteration 396: log likelihood = -100.03748  (backed up)
    
74. Iteration 397: log likelihood = -100.03746  (backed up)
    
75. BFGS stepping has contracted, resetting BFGS Hessian (99)
    
76. Iteration 398: log likelihood =  -100.0372  
    
77. Iteration 399: log likelihood = -100.03715  (backed up)
    
78. Iteration 400: log likelihood = -100.03715  (backed up)
    
79. flat log likelihood encountered, cannot find uphill directi
    
80. > on
    
81. r(430);

复制代码

扫码加我 拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

分享0 收藏0 回帖

我也遇到了这个问题，同样也不知道怎么处理。
我看了看Stata 的说明，里面说可以用过ML function 来建立你自己的Garch 模型，然后通过调整其中的option 来试试看之类的。但是我现在还没有弄明白到底怎么搞。愁的要死

爱吃铜锣烧的诸葛瑾 发表于 2018-11-25 19:05
我也遇到了这个问题，同样也不知道怎么处理。
我看了看Stata 的说明，里面说可以用过ML function 来建立你 ...

好的，谢谢你啊，我找找看。如果实在不行，我可能打算换方法了

Golden-蓉 发表于 2018-11-25 21:06
好的，谢谢你啊，我找找看。如果实在不行，我可能打算换方法了

你好想问下你这个问题解决了吗

爱吃铜锣烧的诸葛瑾 发表于 2018-11-25 19:05
我也遇到了这个问题，同样也不知道怎么处理。
我看了看Stata 的说明，里面说可以用过ML function 来建立你 ...

你好想问下你这个问题解决了吗

1.arch解决极大似然值问题
        a. Help arch
        b.  find maximize option
        c. Technique,help technique,arch RDexpenditure, technique(bfgs) arch(1/1) garch(1/1)可以用
        d.   gtolerance(#)
        (999)可以指定为禁用渐变条件。如果优化器遇到重复的“（备份）”消息，梯度可能仍包含大量值，但无法找到可能性的上坡方向。使用此选项，通常可以获得结果，但是否找到了全局最大似然还不清楚。
        
        当最大化不顺利时，可以将最大迭代次数（参见[R]maximize）设置到优化器似乎卡住的点，并在该点检查估计结果。
        e.     from(init_specs)指定系数的初始值。ARCH模型可能对初始值敏感，并且可能具有与局部极大值相对应的系数值。通过一系列回归得到默认起始值，根据渐近理论，得出的结果对于b和ARMA参数是一致的，对于其余参数通常是合理的。然而，这些值可能并不总是可行的，因为似然函数不能在arch首先选择的初始值处求值。在这种情况下，当ARCH和ARMA参数初始化为零时，重新启动估计函数。有可能，但不太可能，即使是这些值也不可行，而且您必须自己提供初始值。

571269707@qq.co 发表于 2020-11-2 16:16
1.arch解决极大似然值问题
        a. Help arch
        b.  find maximize option

我更改了technique、gtolerance设置，同时，改变了迭代次数

571269707@qq.co 发表于 2020-11-2 16:17
我更改了technique、gtolerance设置，同时，改变了迭代次数

具体怎么更改的啊请问