- Видео 37
- Просмотров 82 235
Paul Sweeting
Добавлен 29 окт 2011
My name is Paul Sweeting, and I'm a Senior Advisor for the Hassana Investment Company. I'm also an honorary Professor of Actuarial Science at the University of Kent. I'm going to use this channel mainly to post videos on pensions, longevity, finance and investment - but if it's newsworthy and even vaguely actuarial, then I might do something here! Enjoy...
Copulas 6.6 - fitting 3D vine copulas
In this video, I show how to fit a 3D vine copula to data, and compare the fit with a multivariate normal distribution. This brings together maximum likelihood estimation, conditional densities, conditional copulas, and provides the building blocks for fitting and assessing any R-vine copula structure
Просмотров: 177
Видео
Copulas 6.5 - densities for 3d vine copulas
Просмотров 1142 месяца назад
Having explored conditional copulas and conditional densities in two dimensions, we can now move onto three-dimensional examples. Understanding these building blocks will allow us to fit vine copulas - first in three dimensions, and ultimately of any size
Copulas 6.4 - conditional densities and conditional copulas
Просмотров 1682 месяца назад
To assess the fit or even parameterise a vine copula, we need to use the probability density function for the data. To make the function more manageable, we need to break this function down. This leads us to conditional densities and conditional copulas - and this video provides the tools for this process.
Copulas 6.3 - deciding vine copula structures
Просмотров 2583 месяца назад
Having described what a vine copula is and how they can look, the next stage is working out how to build one. In this video, I look at a couple of algorithms that can be used to determine the structure of a vine copula, together with an Excel example
Copulas 6.2 - vine copula structures
Просмотров 2483 месяца назад
In this video, I look at the different types of vine copula structure, including how to classify them and the terminology needed to describe them
Copulas 6.1 - an introduction to vine copulas
Просмотров 4113 месяца назад
Vine copulas are a flexible tool for modelling dependence between data series. In this video I give a recap on copulas, explain what vine copulas are and explore the advantages (and disadvantages) relative to "traditional" copulas.
Decision trees 3 - building trees with R
Просмотров 3108 месяцев назад
In this final decision trees video, I show how R can be used to build and prune decision trees, using a dataset of solvent and insolvent companies
Decision trees 2 - constructing trees with continuous data
Просмотров 1688 месяцев назад
In this second video on decision trees, I look at how to create trees using continuous data. Do demonstrate, I construct a tree in Excel for solvent and insolvent firms, based on various accounting ratios. This is quite a laborious way to do it (!) but it shows exactly what is going on when building a decision tree.
Decision trees 1 - introduction
Просмотров 1968 месяцев назад
In this video, I look at the basics of decision trees, including how they're structured and the principles behind working out they're built. For now, I stick to categorical trees, but the principles form the basis of more complex approaches
SVM 7 - support vector machines in R
Просмотров 1658 месяцев назад
Having run through the theory of support vector machines, it's now time to implement an example in R. This video shows how straightforward it is to fit an SVM to credit data, and the impact of changing the fitting parameters.
SVM 6 - the RBF and other kernels
Просмотров 1228 месяцев назад
In SVM 5, I described the kernel method, and why it is useful. But whilst the kernel I used - the quadratic kernel - is great for showing how the kernel method works, isn't so useful for credit modelling. So in this video, I look at the radial basis function (RBF) kernel and show why this it is much more suitable.
SVM 5 - the kernel trick
Просмотров 4599 месяцев назад
The kernel trick is possibly the most important topic related to support vector machines (SVMs). All SVM models in R and Python use it implicitly, but without any real explanation of what it is doing. The kernel trick offers a shortcut to fitting very complex surfaces to separate two groups of data - such as solvent and insolvent firms - using many variables in many forms. The kernel trick is b...
SVM 4 - non-linear support vector machines
Просмотров 1389 месяцев назад
In this fourth SVM video, I talk about what happens when you can't draw a straight line between two data categories - and you need to find a non-linear solution.
SVM 3 - fitting support vector machines using Lagrangians
Просмотров 19010 месяцев назад
This video looks at an alternative approach to fitting support vector machines using the Lagrangian approach. This gives the same answer as the basic optimisation approach, but is essential for employing the "kernel trick", an approach for reducing the computing power needed for more complex optimisations.
SVM 2 - soft margin support vector machines
Просмотров 17010 месяцев назад
In this video, I look at one of the ways to deal with data that is not linearly separable - by using a soft-margin version of the SVM.
SVM 1 - linear support vector machines
Просмотров 28210 месяцев назад
SVM 1 - linear support vector machines
Copulas 5.3 - using R to evaluate and simulate copulas
Просмотров 1,3 тыс.11 месяцев назад
Copulas 5.3 - using R to evaluate and simulate copulas
Copulas 5.2 - using R to fit elliptical copulas
Просмотров 94811 месяцев назад
Copulas 5.2 - using R to fit elliptical copulas
Copulas 5.1 - using R to fit Archimedean Copulas
Просмотров 1,9 тыс.11 месяцев назад
Copulas 5.1 - using R to fit Archimedean Copulas
Using discriminant analysis for credit risk
Просмотров 773Год назад
Using discriminant analysis for credit risk
Credit risk modelling - an introduction
Просмотров 9 тыс.Год назад
Credit risk modelling - an introduction
Copulas 4 - creating simulations with copulas
Просмотров 2,2 тыс.Год назад
Copulas 4 - creating simulations with copulas
Copulas 3.4 - the grand finale (or two more ways to fit copulas)
Просмотров 753Год назад
Copulas 3.4 - the grand finale (or two more ways to fit copulas)
Copulas 3.3 - fitting copulas using maximum likelihood estimation
Просмотров 1,4 тыс.Год назад
Copulas 3.3 - fitting copulas using maximum likelihood estimation
Copulas 3.2 - fitting a copula using method of moments
Просмотров 1,4 тыс.Год назад
Copulas 3.2 - fitting a copula using method of moments
Noise is a bit messed up after minute 49 ..but great explanation thanks
Great Thanks for this clear and amazing explanation lecture!
Great Thanks to you!
One question please: 10:27, 10:50 here what is the "transformed variables"? So they are not the raw observed data? Many Thanks!
Thank you so much for this great lecture! One question here: 3:51, so what you mean is to calc Kendall's tau from data sample, and then convert to corresponding copula's parameter, here alpha, is that right? then what's next step once alpha is obtained? Really appreciate your advice!
Hi Prof Paul, first, want to thank you for great great lecture on this topic. Here I have 2 questions if I may: 1. Starting from normal and then eventually marginal normal, then Choleksy is enough. What does Copula do here? 2. Starting normal end up marginal T-dist, then just need convert normal after Cholesky to T, why we need Copula here? 3. What’s the difference from this video which started with arbitrary random uniform to your later video “Copula fitting”? Whats the diff in terms of work flow? Can we do fitting here too? Thank you so much in advance for your advice!
❤
That's excellent, thank you Paul. Very interesting to see the proximity between the results of Kendalls Tau vs MLE demonstrated too.
Indeed - thanks for all your comments!
Thanks again, Paul. Can’t wait for the next one.
Paul strikes again! Brilliant!
Fantastic. Thank you so much for making this
Hello sir! Thank you for you hard work. Are planning on adding tutorials in R? Regards
Eventually, but there's a bit to get through before that!
fantastic video explaining copulas, thank you! audio quality could be better though :)
at 4:51 is it not the case that any 'traditional' copula can be represented by any regular schema? for instance i refer defn 20 "partial correlation vine specification" of bedford & cooke (2002) for case of multivariate normal. in general the conditional copulas will have complicated form. apologies for my english
What I'm really trying to show is that for a traditional copula, you effectively explicitly define the relationships between all variables, whereas this isn't the case for a vine copula. But there is an absence of mathematical rigour about my explanation!
@@PaulSweeting hmm i dont sure i agree. any vine also allows for explicit pairwise relationship between all variables, although most of pairs are conditioned. or you mean that it is possible to make assumption of conditional independence for pairs in higher order trees? in any case, thank you for the videos!
Really looking forward to move videos on Vine copulas. This was an excellent series BTW, thank you.
Thank you so much Paul. I am looking forward to your explanation of h-functions.
thank you for the video! Insightful 👍.
Sir how to add covariates to these models?
there is somo place like GitHub where I can get the data of this lesson ?
Hi Paul, your videos are incredible. Do you have a video on creating a marginal distribution for a set of returns? I'm reading some papers on using copulas to analyse the dependent structure of equity vs. FX, or gold vs. FX and i can't quite follow their AR-GARCH approaches. Many thanks.
Welcome to Riyadh for the conference. I will join the conference and happy to meet you.
In your slide @13:12, shouldn’t you be showing cdf F instead of pdf f when you have broken down the two distributions?
Congratulations on being elected as IFOA president.
Really great series, thanks you
where can I get the data from
Many Thanks :) The practical aspect to the videos is indeed very helpful. Do you plan to share the data templates used within the videos? That will help in reconciling the results.
Thanks!. But I have some question on copula, how can I ask you?
Promo_SM
Hi sir could you please share the data sets
Hi @Paul Sweeting
Many many thanks for the truly superb videos... please provide the data files as well if possible.
@paulsweeting These videos are really superb! Many thanks for providing them. I was getting overwhelmed with the concept of copulas and these videos were very timely in helping overcoming it.
Around 3 to 4 mins, you may also find the functions in the copula package p2P() and P2p() useful in this regard. p2P() creates a matrix from a given vector of parameters. P2p() creates a numeric vector from a given matrix, currently useful for elliptical copulas.
Really thanks!. But I have some question on the copula generally.
thanks so much for these videos. Would it be possible to get the slides used in the videos after the intro, and also the example dataset showed in this video please?
The data's a bit tricky, but I'm planning on sharing the slides on LinkedIn over the coming weeks - you can find me there at linkedin.com/in/profpaulsweeting
One of the best explanation video channels I saw. I will recommend your channel to my students.
Thanks, much appreciated!
thank you
Excellent presentation and content when fitting your model with xcell, thank you for your time in explaning this and the rest of your videos for those who are neither mathematicians nor actuaries, definetivelly I am subscribing to your channel.
Thank you
Thanks for your video. I want to understand why you do not want a scale of 1 and a mean of 0? (at time 15:07)
I'm trying to simulate (something like) asset returns for the marginal distribution(s). These will have their own expected returns (means) and volatility (=> scale), so when we are converting our uniform outputs (distribution functions) from the simulated copula, we need to convert them to a t distribution with the required location and scale parameters.
I liked your video a lot. However, making a video using commonly used data as a case study will be very helpful.
Noted - just need to make sure that the data in the case study can be shared....
These are the best-ever videos.
Keep up the excellent work! Making a video with some case studies will be a great idea.
Very helpful
Thank you Paul, straightforward explanation.
Thank you... Great video.. gives a good intuitive understanding of Copulas...
why are we only using the columns 2 to 4? and which columns do i use when i have 4 columns in total? (min 14)
Hi! The only reason I used columns 2 to 4 was to show how you can fit a copula to a subset of the data. If you wanted to use just (say) the first 4 columns of 6, it would be retsum_u[,1:4], whereas if you had only 4 columns and you wanted to use them all, it would just be retsum_u (i.e. no brackets needed)
Thanks for the video @@PaulSweeting Is there any way to "check" if our fit is correct by using this colum 1 ?
How do I compare different Archimedian copula models to see which one is better fit if I am using "itau" or "irho" method as there is no Liklihood function?
I'm not sure it's easy in R, as I fitcopula only calculates the likelihood if the approach used is maximum likelihood. You can see this because using (e.g.) the AIC function on the copula object throws up an "NA". You could calculate it directly from the copula density function, but these are quite unwieldy for Archimedean copulas. So probably easiest to just use the maximum likelihood estimation approach if you want to compare likelihoods...
Sir please please do pair wise copula construction and R vines....I am surviving through your videos on your Copulas....
It's on my list...
Great video !, highly didactic.
You are really doing the greater good....Amazing explanation.....🙂