13/12/2019 FRTB and Machine Learning 3: An approach to Non-Modellable Risk Factors using the techniques of Brummelhuis and Luo

elhuis and Luo | LinkedIn

Search Try Premium Free

for 1 Month

FRTB and Machine Learning 3: An

approach to Non-Modellable Risk
Factors using the techniques of
Brummelhuis and Luo
Published on October 21, 2019
David Maher
Market Risk Oversight | Model Validation | FRTB 10 articles Following
| Machine Learning

A major consequence of FRTB Internal Model use will be the impact of Non-Modellable
Risk Factors (NMRF). To be included in the Internal Model, risk factors must pass an
eligibility test, whereby they must consist a sufficient number of "real prices" (see para
31.12 onward). QIS results have shown that the capital impact can be 30%.

The main purpose of this article is to publicize the paper of Brummelhuis and Luo "CDS
Rate Construction Methods by Machine Learning Techniques" (2017), which I believe can
be applied in the FRTB context to further enhance existing proxy methods, reducing the
idiosyncratic basis that would need to be capitalised.

NMRF and Proxys. One approach to reducing capital impact of NMRF's will be via the use
of proxies, whereby there are two well-known approaches (see p10):

1. Simple rules-based approach, ie, use the “closest fit” modellable risk factor is then
used to proxy the non-modellable risk factor.

2. Statistical modelling approach, eg, use regression to look for or manufacture a

modellable risk factor as a proxy for the nonmodellable risk factor. 

In both cases, there is a residual basis that need to be capitalised as a NMRF

Capitalisation of NMRF. From para 33.16: "Capital requirements for each non-modellable
Messaging
risk factor (NMRF) are to be determined using a stress scenario that is calibrated to be at
https://www.linkedin.com/pulse/frtb-machine-learning-3-approach-non-modellable-risk-factors-maher/ 1/5
13/12/2019 FRTB and Machine Learning 3: An approach to Non-Modellable Risk Factors using the techniques of Brummelhuis and Luo | LinkedIn

Search least as prudent as the ES calibration used for modelled risks (ie a loss calibrated to a 97.5%Try Premium Free
for 1 Month
confidence threshold over a period of stress)." So whilst this is not necessarily large, the fact
that the NMRF's are capitalised as a direct add-on (para 33.41) without any diversification to
the internally modeled risks - see para 33.17:

Like Comment Share 16 · 1 Comment

Proxies for NMRF. The section "Principles for the modellability of risk factors that pass the
RFET" (page 78) is worth a careful read. Let us start with Principal two which discerns
between idiosyncratic and general market risk:

Principle two. The data used must allow the

model to pick up both idiosyncratic and
general market risk . . . If the data used in the
model do not reflect either idiosyncratic or
general market risk, the bank must apply an
NMRF charge for those aspects that are not
adequately captured in its model.

So if either of the two proxy methods above were used, the contribution from the proxy
could be incorporated into the ES model (assuing it is modellabel) and the idiosyncratic
basis between the proxy and the "true" risk factor would be capitalised as a NMRF. Better at
least than treating the entire risk factor as non-modellable.

Of most relevance to this article is Principal three:

Principle three. The data used must allow the

model to reflect volatility and correlation of
the risk positions. Banks must ensure that they
do not understate the volatility of an asset (eg
by using inappropriate averaging of data or
proxies).

as well as Principle seven: "The use of proxies must be limited, and proxies must have
sufficiently similar characteristics to the transactions they represent. Proxies must be
appropriate for the region, quality and type of instrument they are intended to represent."
Messaging
https://www.linkedin.com/pulse/frtb-machine-learning-3-approach-non-modellable-risk-factors-maher/ 2/5
13/12/2019 FRTB and Machine Learning 3: An approach to Non-Modellable Risk Factors using the techniques of Brummelhuis and Luo | LinkedIn

Search Brummelhuis and Luo. The authors investigate the use of machine learning techniques in Try Premium Free
for 1 Month
the context of illiquid CDS. The starting point of this paper is to examine the above two
proxy methods:

1. Credit Curve Mapping Approach (Simple rules-based approach). That is, proxy the
CDS by the best choice of name in the Region/Sector/Rating.

2. Cross-sectional Regression Approach (Statistical modelling approach). That is, use

regression to proxy the CDS by Region/Sector/Rating/Seniority.

The authors note that these two techniques do not meet a third condition (EBA 2013), that

"the appropriateness of a proxy spread should

be determined by its volatility and not by its
level".

That is, the above two methods to not incorporate the existing CDS data (albeit illiquid)
which may be more volatile and have other correlation phenomena that they will not take
into account.

The paper is a tour-de-force of ML methods applied to this problem, exploring the

construction using

Linear Gaussian Discriminant Analysis,

Quadratic Discriminant Analysis,

Naive Bayes Classifiers,

k-th Nearest Neighbour,

Logistic Regression,

Decision Trees,

Support Vector Machines (SVM), and

Neural Networks.

Each of these methods is examined in detail for how best they incorporate the additional
data points from the illiquid CDS along with proxys. IMHO a worthy winner of Risk's
Quant Summit Europe Conference Call for Paper 2018.

Application to NMRF. The techniquies of the paper are by no means restricted to illiquid
CDS. CDS could be any asset class. And substituting "real prices" for "illiquid" gives the
application in the NMRF context for FRTB.

So whilst employing a simple rules-based approach or statistical modelling approach may

reduce the NMRF impact by allowing only the idiosyncratic basis to be capitalised this way,
Messaging
https://www.linkedin.com/pulse/frtb-machine-learning-3-approach-non-modellable-risk-factors-maher/ 3/5
13/12/2019 FRTB and Machine Learning 3: An approach to Non-Modellable Risk Factors using the techniques of Brummelhuis and Luo | LinkedIn

Search using one of the methods in Brummelhuis and Luo to incorporate data points consisting of Try Premium Free
for 1 Month
real transactions can help to minimise - or perhaps even eliminate - this basis.

Coursera. For further instruction on the above topics and beyond, one can enroll in the
Coursera course "Fundamentals of Machine Learning in Finance". Of particular interest, this
course has the example of using SVM for the prediction of credit spreads. Graphically, this
is shown in the video from 8:14, where the SVM is compared to an ANN approach, where
the SVM is shown to perform better. I have to admit this was a surprise to me, and has lead
me to turn my attention back to SVM, since I've instead been very focussed on ANN of late.

Another point of interest is the comment at the start of the video (see the transcript) "Here's
a project that I worked on a few years back when I was at JP Morgan". Compare this
to Brummelhaus and Luo "To the best of our knowledge, this paper represents the first
research in the public domain on applications of Machine Learning techniques to the
Shortage of Liquidity Problem for CDS rates." I agree that the paper is the first publicly
published, but it would appear to have been used behind closed doors some time ago. I
would be very interested to hear from anyone involved in this work, as I'm trying to identify
the "Day Zero" when certain ML techniques were first applied in quant finance for a future
article.

Further reading. This article is a sequel to FRTB and Machine Learning, which focussed
on the possible use of ML to speed up computations, as well as the new research in quantile
regression that will enable better analysis on the VaR/ES relationship; and FRTB and
Machine Learning, part 2: Data Management.

#FRTB #MachineLearning

[The views expressed herein are my own and do not reflect those of my employer.]

Report this
Published by
David Maher 10 articles Following
Market Risk Oversight | Model Validation | FRTB | Machine Learning
Published • 2mo
I warmly recommend the paper by Brummelhuis and Luo, whose techniques I believe can be applied to aid with #NMRF in
the #FRTB #machinelearning

Reactions Messaging
https://www.linkedin.com/pulse/frtb-machine-learning-3-approach-non-modellable-risk-factors-maher/ 4/5
13/12/2019 FRTB and Machine Learning 3: An approach to Non-Modellable Risk Factors using the techniques of Brummelhuis and Luo | LinkedIn

Search +8 Try Premium Free

for 1 Month

1 Comment
Add a comment…
Allan Cowan
Managing Director at IHS Markit
Nice article David. Applying the techniques of Brummelhuis and Luo to NMRF is certainly an interesting 3w
idea. The one concern I have had with the Brummelhuis and Luo approach applied to the problem of
proxying illiquid credit spreads for CVA is that it may introduce CVA volatility when the illiquid name
gets classified to a different liquid name. Do you see that as a practical problem?
Like Reply

David Maher
Market Risk Oversight | Model Validation | FRTB | Machine Learning
Following

More from David Maher

FRTB and Machine Learning, part 2: FRTB and Machine Learning Is Machine Learning the new We need to stop talking about
Data Management David Maher on LinkedIn Blockchain? Blockchain
David Maher on LinkedIn David Maher on LinkedIn David Maher on LinkedIn

See all 10 articles

Messaging
https://www.linkedin.com/pulse/frtb-machine-learning-3-approach-non-modellable-risk-factors-maher/ 5/5