0:00

Now I want to give you one example of using the SVM to solve real world

problems.

Here's a project that I worked on a few years back when I was at JP Morgan.

The problem that we were solving back then,

was the problem of modeling credit spreads for illiquid names.

What does this jargon mean?

Well, illiquid names are firms that issue debt or

borrow credit that do not have activated bonds, or

when there are no activated credit default swaps referenced in such a firm.

If you don't know what the Credit Default Swap, or CDS is,

let me briefly explain it.

A CDS is an insurance-like contract between two counter parties that we can

call X and Y that references some company ABC, that can default in the future.

And CDS counter test some standard notional amount

which is equal to $10 million in a US market.

Under a CDS contract that extends for some time T, for example five years.

Counterparty X pays a certain fixed spread, S,

that is some fraction of the notional N to company Y every quarter.

In exchange, counterparty Y agrees to

pay a certain fraction of the contract's notional amount, N.

If the referenced entity, ABC, defaults within the time horizon T.

This fraction is equal to 1 minus R times the notional N of the contract,

where R is a number between 0 and

1 that is called the recovery rate and calculated separately.

The contract then terminates and there are no further payments to counterpart the X.

Otherwise, the payments of Spread to counterpart the X continue till time T.

Therefore, a CDS counter is similar to an insurance

counter to where from X buys protection against default by company ABC, while for

Y itself this protection the company to X for Spread S.

We can see Cashflows in the CDS contract as 2 arc,

sequence of payments that's shown in this graph.

Here, one leg will be payment cashflows is guarantee that raised for

a while by payments of company x to company y.

2:56

CDS only about 500 US names are actively traded in the market.

On the other hand, when we do not have any illiquid CDS that would reference company

ABC, and where there are no bounce of company ABC traded to the market,

in this case, we say that company ABC has a illiquid credit.

It should be estimated from a model From the market.

The main point here is that for illiquid name, there are no market observables that

could be used to find what is called credit spread for

this given credit card counterparty, to which the bank lends money.

Now the bank has a number of borrowers that do not have activated debt,

such as bonds or do not have CDS that would references them.

On the other hand, the bank lends to such firms.

Now, how can the bank calculate its own risk to such firms?

Or in other words,

calculate a spread that should be applied to such borrowers to compensate for

the risk of the event that they would not be able to pay off their debt?

4:18

To this end, we just look at credit spreads of other

similar companies that do have market provided credit spreads.

We can take such market data for about 500 liquidly traded US CDS.

And use it as training data for our regression problem.

More specifically, this provides the outputs of y, so for training data.

But what about the features X in our dataset?

4:49

Well, there is a number of features that are considered important

in driving a credit spreads.

One of them is a company rating given by agency,

such as Moody's or S&P, or internal bank ratings.

Other features include industrial and geographic sector of the issuer.

This is because firms in the same industry or

geography are often exposed to the same common risk factors.

Therefore they tend to default together.

The other feature that goes into prediction is the so-called

Expected Default Frequency, or EDF, that is calculated by Moody's.

5:46

This gives us four predictors, companies' internal rating,

that is computed by the bank, industry sector, geographic sector, and the EDF.

In addition, we had other predictors such as companies' financial ratios or

implied option volatilities, but the basic setting for

the problem includes just these four predictors.

6:09

This offers features for our data set of illiquidly traded CDS

where the output would be observed market spreads for this CDS.

For example, if you want to construct model spreads for

five years horizon, we would use market spreads for CDS.

Just make sure it is five years,

which are the most liquidly traded CDS in the market.

Now it looks like a standard problem for supervised learning.

We have a dataset that we can split into a train and test datasets.

Then we train SVM Regression Model and this data.

In our workload so at the neural network implementation for the same data,

that we could compare with the SVM algorithm adapt.

After such model is built, it can be used in many parts of the banking business.

One of them would be a counterparty risk management.

Such model would provide model based predicted CDS spread for

illiquid counterparties to the bank.

So that the banks could charge them at such spreads to compensate for

its risk of a counterparty default in the future.

And second application is to compute banks capital for created risk in its portfolio.

Similar to the previous case, such calculation also requires knowing

credit spreads of counterparties, you see that from the market or from a model.

7:41

So now we can follow the estimation procedure that we outlined in the previous

video and build such same regression model for credit spread.

We can estimate out of sample errors for such model and compare performance,

its performance with other models, for example, with the neural network model.

It shows out that SVM works a bit better than a neural network on this data.

8:09

Another performance metric for this model can be obtained if we use our model and

constructed proxy to a credit default index, such as CDX.

Because such index is composed of 125

names from different industries and with different ratings.

We can use the outputs of our model to predict the variable of the index.

8:35

In other words, to build a proxy to the index.

And if we do it sequentially, every day or every week,

then we can check how our proxy to the index tracks the index itself.

And the results are shown on these two graphs for

two US graded indexes, CDX, IG, which stands for investment rate.

And CDX and HY, which stands for high yield.