IJ
发帖数: 494 | 1 1.Credit Derivatives Pricing Models: Model, Pricing and Implementation -
Philipp J. Sch?nbucher
2.An Introduction to Credit Risk Modeling (Chapman & Hall/Crc Financial
Mathematics Series) - Christian Bluhm
3.Credit Risk Modelling: The Cutting-edge Collection - Technical Papers
published in Risk 1999-2003
4. Cutting-edge papers on recent Risk Maganize
market |
|
i****e
发帖数: 78 | 2 I guess that you can try few ways
1) reduce the degree of freedom in your model, i.e., reduce model
parameters, use simpler model.
2) or try to calibrate to more market data. for example, if you want
to calibrate the correlation, find some liquid options, which are
sensitive to correlation.
3) or use historical data to determine the undetermined model parameters.
to
some
.g |
|
n****8
发帖数: 23 | 3 If you have both positive or negative bias, it is normal since no model can
predict the rate perfectly. But persistance positive biase means something
wrong.
HW is affine term structure models. If you use 1 factor HW, it doesn't
capture non parallel shift of the term structure. Hence the TS you forecase
could only show paralell movement, which result in persistance positive or
nagative bias.
If you use 2 factor models, it can allow non parallel shift. But there are
many format of 2 factor models |
|
G*****o
发帖数: 1341 | 4 我在一个financial service 公司(starmine) 做 model, 主要是根据wallstreet 各个
投行的analyst 给出的eps estimate 的改变，来做一个monthly 投资的model， 选出
最近得到postive EPS estimate revision。 有谁做过类似的model吗？不了解市场，
也不了解自己做的model 是哪些人在用。欢迎大家讨论。 |
|
b*********e
发帖数: 2 | 5 Please send your resume to j********[email protected]
Job Title
Senior Quantitative Analyst I
Job Description
Using statistical packages such as MatLab, or SAS to build pipeline fallout
models for the purpose of pipeline closing ratio projection and risk
management. This includes retrieving and organizing pipeline data from
database, building statistics models, implementing, validating, documenting
models, and monitoring model performance on a regular basis. In addition,
this candicate will b |
|
g******8
发帖数: 542 | 6 affine jump diffusion model 是heston model上加入compounded poisson jump on
log price and also on volatility. 我看的几篇research paper, 比如“# B.
Eraker. (2003).# Do Stock Prices and Volatility Jump? Reconciling Evidence
from Spot and Option Prices.”，里面就说转换成risk neutral meausre的affine
jump diffusion model的结果is standard in the literature.可是我没找到起源，不
知道是怎么转换的。这里有大牛指点一下么？谢谢！ |
|
C******n
发帖数: 9204 | 7 market肯定是incomplete的，没人能买卖任意数量。
在SV model里是incomplete，但这是this model的缺陷，因为实际上有比该模型更多的
product存在但没考虑。Potentially，任意的shock都可以write contract。正如你所
说，vol是个market，而且liquidity足够好，for the purpose of SV modeling，可以
算是complete。
我主要说的是上面童鞋想模型有点太literal，model说incomplete就能任意设置SDF，
这是不合理的。
to
a
big |
|
k*******i
发帖数: 82 | 8 好吧，我来科普一下吧。 First of all , this is a short rate term structure
model. If you only observe a time series of the short rate r_t, then you can
use MLE or GMM to estimate the model parameters easily because it is an
affine model whose CCF has closed form solution. Now the issue is, you don't
observe the short rate, but the treasury rates only. In principal, the
treasury
rates with different maturities are the functions of the short rate and the
model parameters. Therefore you need to do the followin... 阅读全帖 |
|
a*****a
发帖数: 286 | 9 我是经济学的phd，不是数学出身，快毕业了。论文主要搞the empirical estimation
of stochastic volatility models. 主要用到的方法是bayesian MCMC， 对state
space model, kalman filter or particle filter比较熟悉些。
MSCI’s Equity Risk Modeling Group里的quantitative researcher职位要是面试的
话会主要问哪方面的问题啊？论文做起来都是有深度没广度，想多准备一下职位相关的
知识，希望这里的朋友能给些建设性意见，谢谢。我目前想到的是多看看factor
models和financial time series。 |
|
y*****n
发帖数: 5016 | 10 【 以下文字转载自 Statistics 讨论区 】
发信人: yuxuxin (开水泡妞), 信区: Statistics
标 题: hiring 2 model managers
发信站: BBS 未名空间站 (Fri Aug 1 23:37:10 2014, 美东)
TD bank, Wilmington DE office
will be officially posted on September 1
therefore haven't drafted official job requirements yet
however, below are the key qualifications i am looking for
hands-on modeling experience, preferably acquisition model building
model performance monitoring experience, e.g. creating and reviewing reports
financial industry experience,... 阅读全帖 |
|
t********a
发帖数: 423 | 11 【 以下文字转载自 JobHunting 讨论区 】
发信人: tomorrowba (明天吧), 信区: JobHunting
标 题: Irvine,CA major financial招modeling manager III and director
发信站: BBS 未名空间站 (Wed Apr 15 11:33:05 2015, 美东)
借朋友的ID发帖。
Irvine,CA major financial招modeling manager III and director
需要有predictive modeling, time series modeling experience.Director will
manager a team.
需要有身份。如果有兴趣请email your resume to [email protected]
/* */
谢谢。 |
|
l*******y
发帖数: 22 | 12 你说的正是我所迷惑的地方
hierarchical linear model和random effect model 或者 multi-level model看似很
像，但是有好几个疑点
第一： 前者经常在bayesian学派的书籍中出现，后者几乎只在频率学派的书中出现，
这是否说明这两类模型是两派分别提
出来的？所以还是有所区别的吧。
第二： 有一本书《mixed models》中讲到一点这两者的区别，说他们虽然结构相似，
但是前者的prior是自己给定的，而
后者的“prior”的参数是通过各种优化算法计算出来的，比如em，newton。。
可否指点一下这两者究竟有何关系？ |
|
s**********e
发帖数: 63 | 13 Bayesian 的可以通过matlab 实现：
Step 1: Gibbs Sampling - take marginal probability distribution of your
Estimate Result and run Gibbs Sampler. Starting value - Your OLS result or
whatever model you used to estimate the parameters.
Step 2: Get your posterior and and use Chibs method to yield the Log
Marginal (LME) Likelihood of each model. Take the exponents of the LME, then
you would derive your Marginal likelihood (ML). (Simply taking the
exponents of the LME), Add the ML of all of your models, they s |
|
s*********e
发帖数: 1051 | 14 both response and look-alike models are just buz terms. essentially, they
are logit models. the difference is how to define modeling population and
objectives. |
|
c****s
发帖数: 63 | 15 MSE是用来比较几个Model之间的哪一个MSE的数值最小，如果直接用一个Gamma model的
话，一个MSE的数值恐怕不能说它是好还是不好吧。
老板要用R-square是因为看到一个 R-square就可以判断它是不是好的model了。
我虽不太同意这种做法，但确也找不到其他的这样的test来判断一个GLM Model了
真的是好郁闷呀，不知各位大虾还有没有高招了 |
|
b*******g
发帖数: 513 | 16 为什么在mixed models中加入random effects，response就conditionally
independent了？看了一些书上写的，random effects是用来消除corellated
responses中的correlation的。不管是general linear mixed models,还是genralized
linear mixed models，加入random effects，responses就从correlated，变为
conditionally independent了。基本上书上都这么写，但没写为什么？哪位大虾知道
？还有mixed models中都是什么样的effects能成为random effects？ |
|
f**********t
发帖数: 1001 | 17 1. 想了解一下公司里做这个的时候，是以什么标准来评判一个credit score model的
好坏
的？
2. 另外，一个公司里是不是会有多个credit score model呢？这个问题还望牛人指教
。看相
关的书了解到的情况，是说对历史数据做个logistic regression，这样对于每个
factor都会
有相应的coefficient。然后新的数据到来时，已知coefficient，就能做预测。想问的
是，这
多个credit score model是如何产生的？是不是不同的binning method(相似的组归类)，
factor的取舍，都会产生不同的model?
多谢~~~~ |
|
b*********l
发帖数: 466 | 18 Perform the following ANCOVA models:
Model 1
Dependent variable: Log(BLVALUE)
Independent variables: TREATMENT, AGEYR, SEXSNM
Model 2
Dependent variable: Log(VALUE) Use the value at visit 10
Independent variable: TREATMENT, AGEYR, SEXSNM,
log(BLVALUE) Use the natural log when doing the log transformations
Need the following results reported:
Adjusted mean for Treatment group
Adjusted mean for Placebo group
P-value from the treatment factor in the A... 阅读全帖 |
|
c*****1
发帖数: 131 | 19 主要做marketing modeling (data manipulation, data analysis, build campaign models, model diagnosis and model deployment), need SAS, fresh MS/Ph.D都行，和群，性格好，有潜力,希望本科是学理工科的（数学，统计，CS最好）,愿意relocate到中西部, 2 years contract
Support OPT/H1B
I will not answer questions, please send your resume to s******[email protected] directly if you are interested. We will contact you if you are good candidate.
最好今天,需要这两天确定面世的人选
Thanks a lot, |
|
F******n
发帖数: 160 | 20 Based on your description, I guess you could construct a dynamic model to
account for the "time-dependent" effect. There are standard constructions
for such type of dynamic models, say try google "dynamic linear models."
The specific details for setting up such models depend on your data and
prior/domain knowledge about observations/effects in your studies.
in |
|
s*********e
发帖数: 1051 | 21 it is within a regional bank in cincinnati. if interested, please email your
resume to me w********[email protected].
no relocation package but likely to have some moving assistance. however, we
are very supportive to h1b and green card.
thanks.
1) advanced degree in statistics / econometrics / other quantitative fields
with a solid grasp in modern statistical modeling and data mining and the
completion of statistics courses at the graduate level.
2) high-degree proficiency in SAS base, stat, ets, and en... 阅读全帖 |
|
r****t
发帖数: 276 | 22 诸位大虾，在下试用GENMOD with repeated option去估计一个率，看起来很简单的
model，比如这样：
proc genmod data = dsn;
class subjid y x;
model y = x/link = logit distribution = binomial;
repeated subject = subjid/type=exch;
estimate 'rate a' x 1 0;
run;
可是一旦使用了repeated option 也就是GEE model，出来的率非常邪门。如果直接
proc freq; tables y*x得到的y=1 & x=1的率是70%左右，用GEE model求出来的居然在
30%左右？我注意到用exchangeable option 出来的working correlation〉0.95，这是
不是如此离奇的rate的原因亚？那位大虾对这种情况有经验望不吝赐教，谢先 |
|
h**t
发帖数: 1678 | 23 1) if i save the rf model using the training datset and want to test the rf
model by the test dataset. How does the model was used? it pass the dataset
to all 500 trees(if i use default setting) and give out the everage or it
will pick the one which is most close?
2) rf 的结果是average all trees,我可以看到rules in each tree, 怎么能看到the
final outcome, the average result? 要自己计算吗？
3）能否提供一些rf model 应用的文章
多谢！ |
|
y*****w
发帖数: 1350 | 24 I was running the JM package in R for a joint model for longitudinal and
survival data. I could successfully run the lme() function (for the mixed
model) and the coxph() function (for the Cox model) separately, but when I
ran the JointModel() function, the program returned the following error
message:
Error in fitter(X, Y, strats, offset, init, control, weights = weights, :
Can't fit a Cox model with 0 failures
I double checked my data and there was no any irregular point. What's the
meaning ... 阅读全帖 |
|
a*********y
发帖数: 36 | 25 被问到关于risk model的问题，不太理解，请教版上有相关经验的高人。
1. risk model 用很多credit bureau 的数据，大概两百个吧，但是很多variable 的
missing value 高达90%，这种情况应该如何处理？被告知dataset里面所有的missing
value都被populate 成9999，这样是否可行？如果不可行，应该如何处理？
2. 一般用多少个independent variables 来build model, 一般最后选出来多少个？用
sas run proc logistic 的话，放200最后还是选出来50个左右，怎样可以减少到10-15
个？是不是放得太多？如果不放那么多的话，有什么好的方法可以选择出适量的
variables 放进去 run model?
问题太多，谢谢！ |
|
P*****6
发帖数: 273 | 26 lasso 的使用本版讨论了不少，受益不浅，不过实际应用中还是遇到一个问题，不知高
手如何解决
我的基本问题跟前一段一位老兄相似，就是run不同次结果不同
我的数据只有50个左右，variable 却有400多个
我可以理解可能有多个参数coefficent类似，不过有时参数只有1-2个（可能是最
dominant的那1-2个），有时却有10几个，差别太大，结果中一般来说参数只有1-2个的
CV.err大
另一个解释是，听说glmnet 用coordinate descent 找lamda，可能不是每次都能找到
CV.err最小的
我的问题是：
1。如果我多run几次，选择CV最小的lamda建立的model 业界是否可以接受？我这个
model只是筛选，以后还要验证，所以多引几个variable在model中没有关系
2。使用glmnet有什么技巧可以使建出的model的重复性好些，比如选出的variable都在
10个左右。CV.err也更接近 |
|
w**********y
发帖数: 1691 | 27 我的问题是：
1。如果我多run几次，选择CV最小的lamda建立的model 业界是否可以接受？我这个
model只是筛选，以后还要验证，所以多引几个variable在model中没有关系
--一般不推荐用min而是用lambda.1se吧?特别你的目的只是筛选..
2。使用glmnet有什么技巧可以使建出的model的重复性好些，比如选出的variable都在
10个左右。CV.err也更接近 |
|
t*******t
发帖数: 633 | 28 能做modeling当然做modeling
data management在我看来是最无聊的事情了。
有任何成果反正是modeler做的好，出错很多情况下要怪data有问题。
而且没有高层关心你的data具体是怎么整的。。都关心fancy的model |
|
t********m
发帖数: 939 | 29 多谢回复。经你这么一说，我觉得用survival analysis更合适。现在就是如何建model
的问题。正在看survival analysis using sas这本书，感觉里面的例子似乎没有一个
跟我这个情况一样，很是苦恼。我想将data改成下面这种形式，然后用了163页的model
，这种情况下有些id的age2-age8有missing value，也不知道对不对。
Obs ID MONTHS SEQ EVENT AGE1 AGE2 ..AGE8 BMI1 BMI2...BMI8
1 1 84 8 0 40 41 ....47 27.39999962..
2 2 12 2 1 50 51 ...... 28 27........
proc phreg data=test;
model seq*event(0)=age bmi;
array ag(*) ag1-ag8;
array bm(*) b... 阅读全帖 |
|
w*******9
发帖数: 1433 | 30 I have done a simulation study of the PK model so I know a little bit, but
never did the model fitting.
If you only want the PD parameters, things will be easier. Emax model
assumes
PD = E0 + (Emax-E0)/(1+EC50/Concentration) + error
here (E0, Emax, EC50) is subject specific and follows a joint normal model.
If you want to know more, see
http://www4.stat.ncsu.edu/~davidian/webinar.pdf
nonlinear |
|
b*****y
发帖数: 350 | 31 这问题没那马复杂，
1. 因为C在model里不显著，保留它着不对model有任何贡献。但是，原则上C和B，应该
通过Varclus分析到底应该去掉哪个。
2. 保留C的后果，在multicollinearity存在于model中，对model精度和变量的显著性
检验都有影响。
？？ |
|
c***z
发帖数: 6348 | 32 【 以下文字转载自 JobHunting 讨论区 】
发信人: chaoz (面朝大海，吃碗凉皮), 信区: JobHunting
标 题: Title Source referral: data analyst/modeler
发信站: BBS 未名空间站 (Mon Jun 10 11:14:18 2013, 美东)
Title source is recruiting for modelers. It is a brand new team and I can
send your resume directly to HM.
They sponsor H1B and eventually GC. You need to be able to start working
asap. Previous experience in statistic modeling desired.
My email: n******[email protected]
PS: their job description is outdated, when they say they want analyst, ... 阅读全帖 |
|
p********n
发帖数: 35 | 33 Does anyone know 'Quadratic Model with Plateau'. we want to use this model
to calculate the X threshold where Y begin to decline, with confidence
interval.
The attached is out data. Y=l/[1+A× exp(B×X)], X Y should fit this model.
How to use SAS to graph it and predict after get this model? Who has the
procedure? Thanks a lot.
X Y
0.861 1.000
0.760 0.987
0.680 1.038
0.673 0.976
0.573 1.023
0.452 0.877
0.445 0.805
0.329 0.574
0.221 0.407
0.214 0.211
0.136 0.18... 阅读全帖 |
|
s*********e
发帖数: 1051 | 34 being conservative doesn't mean you can't build a good model.
the real art is how to apply the right tool / model to the best scenario and
still able to tell a good story, which might take some experience.
I did compare all these models off-work and then finally decided that our
team should use fractional logit in lgd model development. |
|
r********n
发帖数: 6979 | 35 我一般情况下都比较喜欢用logistic regression
因为简单，稳定，效果一般也不错
有的时候重要的不是用什么model
重要的是要找到合适的feature
如果feature不好
用什么fancy的model也没用
工作上也试过好几个不同的数据
效果来说
lr， svm的效果都不错， svm尤其是用non-linear kernel的时候也很容易overfit
decision tree效果一般
random forest效果比lr差不多， 不过计算量稍大， 不容易解释
fuzzy logic效果一般， 也很难解释model
nn效果也不错， 不过optimization比较困难， 而且计算量也要大很多， 完全是black
box
最终发现重要的是找到合适的feature
好的feature和差的feature可能可以差30%
不同的model之间的区别可能是10%以内（after all optimization， e.g. feature
selection, imputation， pruning， CV, bootstrapping）
as
to |
|
m******u
发帖数: 11 | 36 一直搞不清楚，大家帮忙看看我理解的对不对？
linear regression model Y=Beta*X+Error X=(x1,x2,...xp)
如果x1...xp的值全是固定的，叫做fixed effect model
如果x1...xp的值全是随机的，叫做random effect model
如果有的固定，有的随机，叫做mixed effect model
是这样吗？ |
|
c****t
发帖数: 19049 | 37 呵呵看来您不知道自己说反了。random effects之所以要用不同的设定是因为所用的
predictor表现特别;而不是因为不同的coefficient设定才把一个predictor变成了
random effect
比如说楼主决定用income作predictor,如果一个intercept,一个slope就可以算出不错
的prediction,那么fixed effect model就搞定。需要做random effect model最简单（
但不是最常见）的情况:不同income有不同的intercepts（假定slope还都一样）。比如
0-50k intercept 1000, 50k-100k intercept 1, 100k+ intercept 10。这样的
relationship不是linear,linear model可以做但predicition不好。而用smoother也许
能提高prediction但不make sense。这时用random effect model效果最好。
为深嘛这predictor这么特别。原因是income其实不是最合适的pr... 阅读全帖 |
|
t*****a
发帖数: 459 | 38 请问楼主的project背景是什么呢？看楼主的描述和so far的讨论，不知会不会分歧在
于做inference的人和做prediction的人思维的差异。做inference的话，非常强调你
model里的每个variable和variable的处理方式都有实际的理由，就是这个variable在
real world里的性质和作用促使你这样处理，而不是因为model fit的原因这样处理。
甚至会经常被问到，你对variable的选择条件和处理方式，是否在做model之前pre
sepcify的。如果是prediction为目的的model，以上这些就不总是那么重要了，但是有
时候也需要考虑到。 |
|
t*****i
发帖数: 11 | 39 【 以下文字转载自 DataSciences 讨论区 】
发信人: tuoxlei (tuo_lei), 信区: DataSciences
标 题: 用10-fold cross-validation 之后怎么挑Model？
发信站: BBS 未名空间站 (Wed May 21 11:29:10 2014, 美东)
一个不大的数据，十几万个record， 一百个变量，用random forest作 binary
classification
因为有over-fitting, 决定用 10-fold cross-validation
做完之后，有十个 random forest Models
下一步 怎么做？
之后 是挑 validation error (on its set-aside 10th hold-out set) 最小的那个
Model吗？（需要一个final model 放进 production system）
Thanks! |
|
c**********e
发帖数: 2007 | 40 When the number of repeated measure is not many, random effect model is good
. I know people use mixed model often in pharmaceutical industry.
But in financial industry, the repeated measure is so many, say unemployment
rate each country in the last 20 years (240 monthly obs). Do we still use
mixed model or random effect model to do panel analysis? Or should we just
use linear regression?
Anybody has experience on this? |
|
h***x
发帖数: 586 | 41 Second this!
We used to build retail models.We multiply the predicted sales value by the
probability the customers would buy our products.
However, the model validation is different (kind of complicated).
use multiplied value to validate two models (one continuous response, one
binary response) simultaneously. If this step fails, you need to rebuild two
models no matter how good they look like individually. |
|
c******d
发帖数: 1150 | 42 matlab我知道的能一个model一个model试，但是需要试的model太多了，想问问有没有
什么软件能够只要输入自变量和因变量就能使用不同model来模拟，算误差的。。
谢谢 |
|
a*****4
发帖数: 986 | 43 This model requests to predict a service provider preference - will assign
each individual with one of the 4 or 5 values which are different brands.
Data given: a survey responder file and a consumer extract file.
Has anyone had experience of this kind of modelling project? It would be
greatly appreciated that you can share some info/comments like modeling
method/process.....
I tried to build 5 1/0 logistic regression models and then score the
consumer extract file. However, how to weight the sc... 阅读全帖 |
|
t********a
发帖数: 423 | 44 【 以下文字转载自 JobHunting 讨论区 】
发信人: tomorrowba (明天吧), 信区: JobHunting
标 题: Irvine,CA major financial招modeling manager III and director
发信站: BBS 未名空间站 (Wed Apr 15 11:33:05 2015, 美东)
借朋友的ID发帖。
Irvine,CA major financial招modeling manager III and director
需要有predictive modeling, time series modeling experience.Director will
manager a team.
需要有身份。如果有兴趣请email your resume to [email protected]
/* */
谢谢。 |
|
m*******e
发帖数: 90 | 45 你说的这个情况属于multicategorical ordinal response的范畴，ANOVA不适用，因为
它是基于linear regression model的。
GLM比较适合，比如可以用baseline category logic model, 也可以用proportional
odds model.
呢？ |
|
s********1
发帖数: 235 | 46 有什么模型能把linear regression model 和 time series model 合起来做
prediction 吗？现在有一些数据，是一堆产品，他们有一些自身的属性的数据，如网
上的ratings 等，他们还有time series 的销售数据，每一种产品有month by month
的销售值，现在要用这些数据做predictive modeling, 想到的方法有对自身属性的数
据可以做linear regression, svm 这样的模型预测，对time series 的销售数据，可
以用 time series 的模型预测，有没有什么模型能把这两种模型结合起来，用一个模
型考虑两方面，进行预测？多谢！ |
|
Y***I
发帖数: 151 | 47 上Data Mining的时候老师提到过Model tree. 就是在Tree的Leaf node上是linear
model或者其他model。
请问什么R package可以这样实现model tree？ CART？
还是需要自己programming？
打算做点小研究，刚刚开始，多谢指教 |
|
m*****g
发帖数: 131 | 48 大佬给我的准备面试的建议。但是看了很多面试的pdf，还是有点蒙。求版上大牛给点
建议。
make sure be prepared with good answers for SAS and any experience about
modeling, and try to be specific about how you create a predictive model
from scratch to implementation
Good answers for SAS 我稍微看了看SAS base和proc sql部分的内容。之前工作也是
做large dataset，涉及到很多很多的merge问题。准备就说这些，其他好像也没有太多
内容。还有部分data cleansing什么的。
主要问题在how you create a predictive model from scratch to implementation.
楼主之前工作做的都是market risk，基本都是用simulation在做。而且model很多都受
限于大量的constrains 和 assumpt... 阅读全帖 |
|
t**********y
发帖数: 374 | 49 I did some test runs using built-in dataset and got different outputs:
1. ===
> test=lmer(Reaction~Days+(Days|Subject),sleepstudy)
> summary(test)
Linear mixed model fit by REML ['lmerMod']
Formula: Reaction ~ Days + (Days | Subject)
Data: sleepstudy
REML criterion at convergence: 1743.6
Scaled residuals:
Min 1Q Median 3Q Max
-3.9536 -0.4634 0.0231 0.4634 5.1793
Random effects:
Groups Name Variance Std.Dev. Corr
Subject (Intercept) 612.09 24.740
... 阅读全帖 |
|
l********i
发帖数: 231 | 50 刚开始学习做look alike model
软件是SPSS modeler
但是没有找到资料参考，请问哪里能查到做这个model的方法呢？不需要是spss
modeler软件的
先谢了 |
|
