j*****e
发帖数: 333 | 1 目前predict volatility的方法都有什么？公认最好的是什么？我印象中是有用
historical vol, GARCh, realized vol, implied vol, VIX index这几种的是么？还
有其他什么办法？大概是怎么用的呢？ |
l******i
发帖数: 1404 | |
j*****e
发帖数: 333 | 3 我也不喜欢garch，historical vol也不好用，你能不能说说Implied vol and VIX?
VIX是不是普遍被认为最好用的predictor? 如果发现一个比它更好的predictor，是不
是就是新发现了？
【在 l******i 的大作中提到】
: 以我的经验Garch是骗人的，但大家都要用。
|
T*******t
发帖数: 9274 | 4 毛估估
【在 j*****e 的大作中提到】
: 目前predict volatility的方法都有什么？公认最好的是什么？我印象中是有用
: historical vol, GARCh, realized vol, implied vol, VIX index这几种的是么？还
: 有其他什么办法？大概是怎么用的呢？
|
z****g
发帖数: 1978 | 5 intraday and daily are complete different. Daily is easier, intraday is hard |
w**********y
发帖数: 1691 | 6 In academia, Levy driven O-U process is a popular one.
比如这篇文章:
http://www.kellogg.northwestern.edu/faculty/todorov/htm/papers/pvi.pdf
他给出了GMM的model calibration方法..我对此方法的适用性画一个大大的问号
他另外还有很多文章关于这个方面的.
我dissertation做这个方向,我的理解,大体思路是..先给一个unbiased estimation of
integrated volatility. 然后用一个mean reversion的model去modelling
volatility. |
w**********y
发帖数: 1691 | 7 然后结合了option做的,我印象中Carr和Oomen都有相关的paper.. |
w**********y
发帖数: 1691 | 8 你好像在chicago..那Uchicago的Mykland和UIC的 Zhang Lan都做相关的research..
(btw,这是夫妻店啊..我太八卦了) |
k****o
发帖数: 11 | 9 好像看过一个用Normal Inverse Gaussian来描述vol future的paper。 |
j*****e
发帖数: 333 | 10 呵呵，他俩确实是夫妻店，但这些文章大家都知道主要是夫做出来的，妻用Mykland的话说就是：”she is good at having a big picture"，哈哈。Mykland确实很聪明。但是他做的几片有名文章主要是关于怎么处理microstructure noise的，主要是关于怎么measure volatility, mainly RV，没怎么做预测好像？Mykland是张兰的博士时导师。我是想看predict方面哪些做的比较好。我看的几片文章貌似都推荐说VIX的预测最好。你说的OU process Viktor好像有提到说它比较能describe some patterns in real data，但还没做预测方面的研究？
【在 w**********y 的大作中提到】
: 你好像在chicago..那Uchicago的Mykland和UIC的 Zhang Lan都做相关的research..
: (btw,这是夫妻店啊..我太八卦了)
|
|
|
j*****e
发帖数: 333 | 11 什么是毛咕咕？
【在 T*******t 的大作中提到】
: 毛估估
|
z****g
发帖数: 1978 | 12 杭州人？
【在 T*******t 的大作中提到】
: 毛估估
|
z****g
发帖数: 1978 | 13 杭州人？
【在 T*******t 的大作中提到】
: 毛估估
|
p******i
发帖数: 1358 | 14 mykland 思维实在跳跃的太快了
一般人类很难跟上。。。
的话说就
是：”she is good at having a big picture"，哈哈。Mykland确实很聪明。但是他
做的几
片有名文章主要是关于怎么处理microstructure noise的，主要是关于怎么measure
volatility, mainly RV，没怎么做预测好像？Mykland是张兰的博士时导师。我是想看
predict
方面哪些做的比较好。我看的几片文章貌似都推荐说VIX的预测最好。你说的OU
process Viktor好
像有提到说它比较能describe some patterns in real data，但还没做预测方面的研
究？
【在 j*****e 的大作中提到】
: 呵呵，他俩确实是夫妻店，但这些文章大家都知道主要是夫做出来的，妻用Mykland的话说就是：”she is good at having a big picture"，哈哈。Mykland确实很聪明。但是他做的几片有名文章主要是关于怎么处理microstructure noise的，主要是关于怎么measure volatility, mainly RV，没怎么做预测好像？Mykland是张兰的博士时导师。我是想看predict方面哪些做的比较好。我看的几片文章貌似都推荐说VIX的预测最好。你说的OU process Viktor好像有提到说它比较能describe some patterns in real data，但还没做预测方面的研究？
|
j*****e
发帖数: 333 | 15 是么？这大概是为什么他课教的很不好，学生听不懂，考试题难的他不知道为什么学生
分数那么低。。。
【在 p******i 的大作中提到】
: mykland 思维实在跳跃的太快了
: 一般人类很难跟上。。。
:
: 的话说就
: 是：”she is good at having a big picture"，哈哈。Mykland确实很聪明。但是他
: 做的几
: 片有名文章主要是关于怎么处理microstructure noise的，主要是关于怎么measure
: volatility, mainly RV，没怎么做预测好像？Mykland是张兰的博士时导师。我是想看
: predict
: 方面哪些做的比较好。我看的几片文章貌似都推荐说VIX的预测最好。你说的OU
|
m*******r
发帖数: 98 | 16 Instead of predicting vol by assuming some underlying process, you can
simply run regression of vol index to lag variables. MIDAS might be an
option.
【在 j*****e 的大作中提到】
: 目前predict volatility的方法都有什么？公认最好的是什么？我印象中是有用
: historical vol, GARCh, realized vol, implied vol, VIX index这几种的是么？还
: 有其他什么办法？大概是怎么用的呢？
|
j*****e
发帖数: 333 | 17 this is exactly what I am thinking abt. but to check if the prediction is good or not, people use Realized vol to proxy the true vol, or say the one corrected for microstructure noise/neweywester errors, right? and people think VIX index is the best to predict vol, right? I constructed a vol predictor that has a much higher R^2 than VIX index on predicting vol (proxied by realized vol). of course I need further more test and robust check to confirm mine really works better than VIX. So I also wonder if there is any better vol predictor than VIX that already exist for now?
btw, what is MIDAS? thanks
【在 m*******r 的大作中提到】
: Instead of predicting vol by assuming some underlying process, you can
: simply run regression of vol index to lag variables. MIDAS might be an
: option.
|
w**********y
发帖数: 1691 | 18 Google MIDAS + volatility.
It is an interesting method, which is more statistical or machine learning
style.
And it has kind of similar flavor as we did in the 2010 INFORMS Data Mining
Contest.
@jeannie:
I am confused with ur statement.
VIX is calculated on a combination of different stocks from SPX. You can
find the detailed formula on CBOE. I can't understand why it could be a good
predictor for a single stock.
What is ur purpose?
option pricing?
Volatility swap pricing?
volatility arbitrage? |
N******r
发帖数: 642 | 19 linear regression dont mean nothing because vol is not supposed to be linear
. my 2%. |
j*****e
发帖数: 333 | 20 yes, vol is not supposed to be linear, but what they are talking abt is time
series modeling, not simply linear regression on some x variables, right?
linear
【在 N******r 的大作中提到】
: linear regression dont mean nothing because vol is not supposed to be linear
: . my 2%.
|
|
|
j*****e
发帖数: 333 | 21 Hi weekendsunny, thank you for the info. I checked the paper you gave, it seems to me that you are only comparing with GARCH, right? Garch is not good compared with VIX and other measures which has been reported in a paper in 2005 or even earlier in a paper in 2001 if my memory is precise. Or you are doing a different thing? thanks.
btw, vix is a measure for market volatility. Vix is the just the model free implied vol constructed for S&P500. In the literature, this model free implied vol is still reported to work well for both market or individual options. Using the same formula, you should be able to construct the vol for individual stocks (if they have the corresponding data). My construct can be used for both S&P500 or individual stocks. it is non parametric.
Mining
good
【在 w**********y 的大作中提到】
: Google MIDAS + volatility.
: It is an interesting method, which is more statistical or machine learning
: style.
: And it has kind of similar flavor as we did in the 2010 INFORMS Data Mining
: Contest.
: @jeannie:
: I am confused with ur statement.
: VIX is calculated on a combination of different stocks from SPX. You can
: find the detailed formula on CBOE. I can't understand why it could be a good
: predictor for a single stock.
|
m*******r
发帖数: 98 | 22 yeap. MIDAS is basically a nonlinear regression.
Actually, if you have some vol index that you want to predict, then it is a
standard supervised machine learning problem. You can simply throw the data to
a support vector machine solver with RKB.
You can do feature selection and model selection as well.
time
【在 j*****e 的大作中提到】
: yes, vol is not supposed to be linear, but what they are talking abt is time
: series modeling, not simply linear regression on some x variables, right?
:
: linear
|
j*****e
发帖数: 333 | 23 i see. it is a way/model for forecast. so basically we can throw everything,
such as implied vol, or vix, or historical vol, etc, into this thing to
forecast?
a
to
do
【在 m*******r 的大作中提到】
: yeap. MIDAS is basically a nonlinear regression.
: Actually, if you have some vol index that you want to predict, then it is a
: standard supervised machine learning problem. You can simply throw the data to
: a support vector machine solver with RKB.
: You can do feature selection and model selection as well.
:
: time
|
w**********y
发帖数: 1691 | 24 As I know, both Adaboost and SVM or kernel SVM can't do feature selection.
You can only control some (tune) parameters, like the lambda in gaussian
kernel. or number of iterations in Adaboost.
Surely, you can do dimension reduction first.
In addition, SVM and Adaboost are classification methods. They only give 0
or 1 predictions (or probability of 1), not continuous predictions.
My research is in estimations of realized volatility/integrated volatility,
based on high frequency data with microstructure noise and jumps. It is on
the same direction as Zhang Lan and Mykland, Todorov, Jacod and Sahalia's
work.
I am not familiar with volatility prediction, beyond the model based method..
Need to sleep now, talk with you guys tomorrow. nite.
a
to
do
【在 m*******r 的大作中提到】
: yeap. MIDAS is basically a nonlinear regression.
: Actually, if you have some vol index that you want to predict, then it is a
: standard supervised machine learning problem. You can simply throw the data to
: a support vector machine solver with RKB.
: You can do feature selection and model selection as well.
:
: time
|
m*******r
发帖数: 98 | 25 Sorry, I was thinking about another problem when I mentioned SVM and Boost.
SVM can be extended to continuous variables after taking care of the
boundary sensitivity. Not sure about Boost. I have modified the post.
,
【在 w**********y 的大作中提到】
: As I know, both Adaboost and SVM or kernel SVM can't do feature selection.
: You can only control some (tune) parameters, like the lambda in gaussian
: kernel. or number of iterations in Adaboost.
: Surely, you can do dimension reduction first.
: In addition, SVM and Adaboost are classification methods. They only give 0
: or 1 predictions (or probability of 1), not continuous predictions.
: My research is in estimations of realized volatility/integrated volatility,
: based on high frequency data with microstructure noise and jumps. It is on
: the same direction as Zhang Lan and Mykland, Todorov, Jacod and Sahalia's
: work.
|
