|
1) Market
Risk: Build
Confidence from your Confidence.
Horizon and
Confidence Intervals should always make sense. This sounds trivial, but
from experience many practitioners are still at odds as to why this is
so important.
A Confidence
level is never chosen on the basis that higher is better.
In many
cases it's the opposite!
For market risk, the Confidence level should be chosen based on the sampling frequency
of the revaluations and
the time required to observe excession and thus validate the model.
For example, a Portfolio Manager who seeks 99% confidence and rebalances his portfolio monthly will,
in theory, need
100 Months (8 1/4 years!!) to get his first error and 1 Year and 8
months if he used 95% confidence!
With Credit risk models the Confidence level is often selected by
taking the complement of the default probability of the part.
(i.e. a party with 0.2% default over 1 year would assume 99.8%
confidence)
2) Know thy
errors.
Understand
where your errors are coming from !
You should have a clear escalation procedure established for each
excession!
Excession should be used proactively to tweak and improve your models and methodologies.
As mentioned above, there is a close relationship between errors
and the granularity of your results.
A common mistake is to
assume the more samples or simulations the better.
This is never the
case in practice !
Find
out your optimal number of runs rather than running simulations in
the dark!
3) Report in
line with your line of Business.
Only use the
reports that make sense to your line of business.
Incremental VaR is only usefull at the individual trader level.
A CEO has no
need for Marginal or Incremental VaR.
On the other hand counterparty / concentration risks are vital
in many aspects involved with his role.
4) Never rely
on Proprietary Models ALONE.
Many examples
abound, from convertible deals that were mispriced due to dividends or badly valued interest rate
derivatives.
If you specialize in a type of product and you believe in your
model, use it!
But always make sure you complement it with an
industry standard equivalent (i.e. static replication).
5)
Bend your system until it breaks!
To
ensure correct pricing and sensitivity of your most complex deals, create two portfolios. The first should hold your
complex product(s), the other should hold a portfolio that
statically replicates your position's payout in the opposite
direction (long for short, short for long).
The total should always be 0 or at least very close to 0 and the individual
aggregations should not ! If this is not the case there's something
terribly wrong with your
engine!.
i.e. Reverse
Convertible - Vs- Bond & put on Equity.
paying swaption - receiving swaption - Vs Fixed Float Bond
Forward, Swaps as FRAs, etc
6) Be Sharp
about your returns !
You don't
need to be an asset manager to use a Benchmark to Measure
Performance!
Use a proxy to Measure your own performance and
shortfall!
7) Use
Relative Measures to Relate to Others.
Relative
Percentage VaR (i.e. VaR dividend by NPV) presents a clear
advantage over VaR in monetary terms:
It can be readily compared to other
levels and thus help to identify so called hot-spots !
On the other hand, VaR in monetary terms is useful when fixing limits and
requesting collateral!
8) Methodologies
are complementary:
Never assume one
methodology is better than another.
Parametric VaR
should be used to rapidly identify obvious problems.
Historical Simulation has
the desirable property of maintaining specific risk (and the
drawback of clustering!)
9) Always
complement with stress:
Every analysis should be complemented with stress tests.
Ideally you should hold a library of stress tests, including
past crashes, extreme events, industry and correlation breakdowns.
10) Seek glamour
in Data!
It is not
glamorous, it tends to be extremely reactive, but market data is without a
doubt the most important aspect of Risk Measurement. Without
data there cannot be any measurement.
Erroneous data only produces
misleading results.
10.1 If data
is missing, find a proxy!
If data is missing for a series, or the price is stale because it
hasn't moved for some time, you might want to consider a proxy:
A proxy should mimic closely the asset itself or should share at
least some basic characteristics.
The recognition of Credit synthetic asset based models by regulatory
authorities should dispell any doubts regarding proxies.
If data is missing proxy it ! Either from a series of country
/sector indices or other data.
Yield Curves are to Market what Credit Curves are to credit
Choose only the most liquid vertices. Do not use 15 vertices if
there are only 8 that are highly liquid!
Some practitioners
drop quirky correlations by imposing 0 (thereby summing variance: a compromise between +1 and
-1!).
others might replace it by a mean, while other will accept any
change coming from a trusted system.
A note on
Interest Rate Books and Correlation breakdowns. Correlations tend to
breakdown. Most people assume short
interest rates MUST be
highly correlated. This is not always the case. You only need a few
basis point movements on one vertex and the opposite movements (with
perhaps different magnitude) on another to see correlation fall from
95% to 45%. (especially with decay).
Solution Keep a
history of Correlations.
10.2 Never Mix Par rates
& Zero
Coupon Prices.
Par Rates and Zero Coupon Prices offer alternate routes to solving
the same problem. If day count conventions and calculation are kept
separate they yield almost the same results.
However, Zero coupon Prices are highly sensitive to changes in prices
in the term structure, a single basis point drop in a one month
vertex and a one basis point upward move in the 10 year vertex can
trigger a collapse in the correlation between the two vertices.
On the other hand, par rates / prices are the aggregation of
different maturities, small changes in the term structure do not
create such large swings. A common mistake, due to the complexity and the sheer number of
instruments involved consists in mixing the two. This can lead to
serious mistakes in risk measurement over long periods of time.
If you are using both par rate and zero coupon prices, then clues pointing to
this problem might appear when there are twists in the term structure of
interest rates or when rates change direction.
10.3
Compute Relative Multiplicative Spreads.
Spreads volatilities are
notiriously tricky. If you transform risky rates into additive
spreads
you might get misleading results. Instead use relative or
multiplicative spreads. Relative spreads
can be combined and thus computed via
substraction and addition!
|
