Chapter 2: Non-Parametric Approaches To apply the bootstrap, we create a large num

ber of new samples, each observation of which

the non-parametric approaches, which seek to
is obtained by drawing at random from our
estimate risk measures without making strong
original sample and replacing the observation
assumptions about the relevant (e.g., P/L)
after it has been drawn. Each new 'resampled'
distribution.
sam ple gives us a new VaR estimate, and we
2.1 COMPILING HISTORICAL SIMULATION DATA: can take our 'best' esti mate to be the mean of
these resample-based estimates. The same
The first task is to assemble a suitable P/L series approach can also be used to produce resample-
for our portfo lio, and this requires a set of based ES estimates-each one of which would be
historical P/L or return observations on the the average of the losses in each resample
positions in our current portfolio. exceeding the resample VaR-and our 'best' ES
estimate would be the mean of these estimates.

Historical Simulation Using Non-parametric

Density Estimation

To appreciate what this involves, we must

recognize that basic HS does not make the best
use of our information.

Non-parametric density estimation also allows

2.2 ESTIMATION OF HISTORICAL SIMULATION us to estimate VaRs and ESs for any confidence
VAR AND ES levels we like and so avoid constraints imposed
by the size of our data set and treats our data as
Basic Historical Simulation: if they were drawings from some unspecified or
Having obtained our historical simulation P/L unknown empirical distribution function
data, we can estimate VaR by plotting the P/L (or drawing straight lines through the mid-points of
L/P) on a simple histogram and then reading off the tops of histogram bars is not the best we can
the VaR from the histogram. If we take our do: we could draw smooth curves that meet up
confidence level to be 95%, our VaR is given by nicely, and so on. This is exactly the point of
the x-value that cuts off the upper 5% of very non-parametric density estimation, the purpose
high losses from the rest of the distribution. of which is to give us some guidance on how
Given 1000 observations, we can take this value 'best' to draw lines through the data points we
(i.e., our VaR) to be the 51 st highest loss value, have.
or 1. 704. 3 The ES is then the aver age of the 50
highest losses, or 2.196. So although kernel methods are better in theory,
they do not necessarily produce much better
Bootstrapped Historical Simulation estimates in practice. There are also practical
a bootstrap procedure involves resampling from reasons why we might prefer simpler non-
our existing data set with replacement. A parametric density estimation methods over
bootstrapped estimate will often be more kernel ones. Although the kernel methods are
accurate than a 'raw' sample estimate, and theoretically better, crude methods like drawing
bootstraps are also useful for gauging the straight-line 'curves' through the tops of
precision of our estimates. histograms are more transparent and easier to
check.
2.3 ESTIMATING CONFIDENCE INTERVALS FOR One such approach is to weight the relative
HISTORICAL SIMULATION VAR AND ES importance, of our observations by their age.

The methods considered so far are good for

giving point estimates of VaR or ES, but they
don't give us any indication of the precision of
these estimates or any indication of VaR or ES.
there are methods to produce confidence
intervals for our risk estimates.

An Order Statistics Approach to the Estimation

of Confidence Intervals for HS VaR and ES

One of the most promising methods is to apply

the theory of order statistics, this approach
gives us, not just a VaR (or ES) estimate, but a
For example, if we are using a spreadsheet, we
complete VaR (or ES) distribution function from
can order our P/L observations in one column,
which we can read off the VaR (or ES)
put their weights w(i) in the next column, and go
confidence interval.
down that column until we reach our desired
'hsvarpdfperc' function percentile. Our VaR is then the negative of the
corresponding value in the first column. And if
A Bootstrap Approach to the Estimation of
our desired percentile falls between two
Confidence Intervals for HS VaR and ES
percentiles, we can take our VaR to be the
We can also estimate confidence intervals using (negative of the) interpolated value of the
a boot strap approach: we produce a corresponding first-column observations.
bootstrapped histogram of resample-based VaR
This not only means that age-weighted VaR
(or ES) estimates, and then read the confidence
estimates are more responsive to large losses,
interval from the quantiles of this histogram.
but also makes them better at handling clusters
Using the basic percentile interval approach of large losses.
outlined in Appendix 2 to this chapter, the 90%
Volatility-weighted Historical Simulation
confidence interval for our VaR is [1.554, 1. 797].
We can also weight our data by volatility.
2.4 WEIGHTED HISTORICAL SIMULATION

In short, our HS P/L series is constructed in a

way that gives any observation the same weight
on P/L provided it is less than n periods old, and
no weight (i.e., a zero weight) if it is older than
that. This weighting structure has a number of
problems: it can overestimate or underestimate
the true risks can also make risk estimates
unresponsive to major events.

Age-weighted Historical Simulation

For instance, we can combine it with the age- number of drawings from this data set, which
weighted approach if we wished to increase the we now treat as a sample, replacing each one
sensitivity of risk estimates to large losses, and after it has been drawn, and multiply each
to reduce the potential for distortions and ghost random drawing by the AGARCH forecast of
effects. We can also combine the HW approach tomorrow's volatility. If we take M drawings, we
with OS or bootstrap methods to estimate therefore get M simulated returns, each of
confidence intervals for our VaR or ES. which reflects current market conditions
because it is scaled by today's forecast of
Correlation-weighted Historical Simulation
tomorrow's volatility. Finally, each of these
We can also adjust our historical returns to simulated returns gives us a possible end-of-
reflect changes between historical and current tomorrow portfolio value, and a corresponding
correlations. possible loss, and we take the VaR to be the loss
corresponding to our chosen confidence level.

2.5 ADVANTAGES AND DISADVANTAGES OF

NON-PARAMETRIC METHODS
*The example is not well explained in the book
Advantages:
Filtered Historical Simulation
Easy to understand and apply without requiring
This is a form of semi-parametric bootstrap complex assumptions.
which aims to combine the benefits of HS with
Can accommodate fat tails, skewness, and other
the power and flexibility of conditional volatility
irregularities in data that parametric methods
models such as GARCH. It does so by
struggle with.
bootstrapping returns within a conditional
Can be used for any type of financial positions,
including derivatives.

Can be implemented using common tools like

spreadsheets, making it accessible.

Widely Used in Practice

Disadvantages

the risk estimates might not accurately reflect

volatility (e.g., GARCH).
the current situation.
The second step is to use the model to forecast
Non-parametric methods struggle to account for
volatility for each of the days in a sample period.
extreme, unlikely events that didn’t happen in
These volatility forecasts are then divided into
the historical data but are still possible.
the realised returns to produce a set of
standardised returns. These standardised
returns should be independently and identically
distributed (iid), and therefore be suitable for
HS. Assuming a 1-day VaR holding period, the
third stage involves bootstrapping from our data
set of standardised returns: we take a large