Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
Quantitative Risk Management: Lecture 11 Non-parametric Methods 1
Tim Bailey
Nottingham University Business School
1/13
Outline
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
Non-Parametric Methods
2/13
Outline
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
Non-Parametric Methods
Non-Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
2/13
Outline
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
Non-Parametric Methods
Non-Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
2/13
Non-Parametric Statistics
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
3/13
Non-Parametric Statistics
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
No assumption made regarding what specific model generates the data
3/13
Non-Parametric Statistics
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
No assumption made regarding what specific model generates the data
Hence no parametric model to fit to data
3/13
Non-Parametric Statistics
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
No assumption made regarding what specific model generates the data
Hence no parametric model to fit to data
Many variants of Non-parametric Methods –
smoothing (regression, density estimation) – illustrative examples shortly.
3/13
Non-Parametric Statistics
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
No assumption made regarding what specific model generates the data
Hence no parametric model to fit to data
Many variants of Non-parametric Methods –
smoothing (regression, density estimation) –
illustrative examples shortly.
We focus on Non-Parametric Simulation – specifically Historical Simulation:
3/13
Non-Parametric Statistics
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
No assumption made regarding what specific model generates the data
Hence no parametric model to fit to data
Many variants of Non-parametric Methods –
smoothing (regression, density estimation) –
illustrative examples shortly.
We focus on Non-Parametric Simulation – specifically Historical Simulation:
Basic Historical Simulation (Single asset, two assets)
3/13
Non-Parametric Statistics
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
No assumption made regarding what specific model generates the data
Hence no parametric model to fit to data
Many variants of Non-parametric Methods –
smoothing (regression, density estimation) –
illustrative examples shortly.
We focus on Non-Parametric Simulation – specifically Historical Simulation:
Basic Historical Simulation (Single asset, two assets)
(and a bit on bootstrapping)
3/13
Non-Parametric Regression
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
Linear regression shown in red Nonparametric regression
●●
●
●
●
●
●●
●●● ●●
●
●
●
●● ●
●
●
●
● ●●●
●
●
●
2345
Car Weight
●
●●
4/13
Miles Per Gallon
10 15 20 25 30
Non-Parametric Regression
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
Linear regression shown in red
Non-parametric regression in blue.
Nonparametric regression
●●
●
●
●
●
●●
●●● ●●
●
●
●
●● ●
●
●
●
● ●●●
●
●
●
2345
Car Weight
●
●●
4/13
Miles Per Gallon
10 15 20 25 30
Non-Parametric Density Estimation
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
5/13
Non-Parametric Density Estimation
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
Could the underlying density ‘really’ be represented by a smoother relationship?
5/13
Non-Parametric Density Estimation
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
Could the underlying density ‘really’ be represented by a smoother relationship?
Leads to idea of nonparametric density estimation – kernel based estimates.
5/13
Non-Parametric Density Estimation
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
Could the underlying density ‘really’ be represented by a smoother relationship?
Leads to idea of nonparametric density estimation – kernel based estimates.
Here, estimated density is the vertical sum of 6 identical gaussians. [Source Wiki!]
5 / 13
Historical Simulation: Single Asset
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
6/13
Historical Simulation: Single Asset
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
From Lec 8 we have the basic mapping from returns to Loss:
Loss = −Pt × Rt
ie the loss is the negative of the current portfolio value times the returns.
6/13
Historical Simulation: Single Asset
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
From Lec 8 we have the basic mapping from returns to Loss:
Loss = −Pt × Rt
ie the loss is the negative of the current portfolio value times the returns.
With HS we use the historical returns of the asset for Rt.
Historical Returns Rt
Predicted Losses
-Pt x Rt
(Pt = 1000)
6/13
Historical Simulation: Single Asset
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
From Lec 8 we have the basic mapping from returns to Loss:
Loss = −Pt × Rt
ie the loss is the negative of the current portfolio value times the returns.
With HS we use the historical returns of the asset for Rt.
Historical Returns Rt
Predicted Losses
-Pt x Rt
(Pt = 1000)
6/13
Historical Simulation: Single Asset
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
From Lec 8 we have the basic mapping from returns to Loss:
Loss = −Pt × Rt
ie the loss is the negative of the current portfolio value times the returns.
With HS we use the historical returns of the asset for Rt.
Historical Returns Rt
Predicted Losses
-Pt x Rt
(Pt = 1000)
6/13
Historical Simulation: Single Asset
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
From Lec 8 we have the basic mapping from returns to Loss:
Loss = −Pt × Rt
ie the loss is the negative of the current portfolio value times the returns.
With HS we use the historical returns of the asset for Rt.
Historical Returns Rt
Predicted Losses
-Pt x Rt
(Pt = 1000)
6/13
Historical Simulation: Single Asset
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
From Lec 8 we have the basic mapping from returns to Loss:
Loss = −Pt × Rt
ie the loss is the negative of the current portfolio value times the returns.
With HS we use the historical returns of the asset for Rt.
Historical Returns Rt
Predicted Losses
-Pt x Rt
(Pt = 1000)
6/13
Historical Simulation
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
With the resulting discrete set of losses, we can measure risk as per Lecture 4.
Histogram of DAX portfolio losses
Frequency
0 200 400 600 800
−50 0 50 100
d.loss
VaR95 = 15.92 ES95 = 23.99
7/13
Historical Simulation: Two Assets
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
8/13
Historical Simulation: Two Assets
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
From Lec 5 we have the basic mapping from returns to Loss:
Loss=−P ×( ×R + ×R ) t A At B Bt
8/13
Historical Simulation: Two Assets
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
From Lec 5 we have the basic mapping from returns to Loss:
Loss = −Pt × (A × RAt + B × RBt)
With HS we use the historical returns of both
assets for R
At
and R
Bt
.
Historical Returns Predicted Losses Dax FTSE
-Pt x (wA x RAt + wB x RBt)
Using:
Pt = 1000 wA = 0.75 wB = 0.25
8/13
Historical Simulation: Two Assets
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
From Lec 5 we have the basic mapping from returns to Loss:
Loss = −Pt × (A × RAt + B × RBt)
With HS we use the historical returns of both
assets for R
At
and R
Bt
.
Historical Returns Predicted Losses Dax FTSE
-Pt x (wA x RAt + wB x RBt)
Using:
Pt = 1000 wA = 0.75 wB = 0.25
8/13
Historical Simulation
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
With the resulting discrete set of losses, we can again measure risk as per Lecture 4.
Histogram of DAX/FTSE Portfolio losses
Frequency
0 200 400 600 800
−40 −20 0 20 40 60 80
df.loss
VaR95 = 14.23 ES95 = 19.31
9/13
Historical Simulation
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
10/13
Historical Simulation
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
Benefits of approach
10/13
Historical Simulation
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
Benefits of approach
Letting data drive estimation
10/13
Historical Simulation
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
Benefits of approach
Letting data drive estimation
No (false) distributional assumptions about returns.
10/13
Historical Simulation
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
Benefits of approach
Letting data drive estimation
No (false) distributional assumptions about returns.
Outlying (dangerous) returns are represented in loss distribution, whereas they might be ignored
with parametric estimation.
10/13
Historical Simulation
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
Benefits of approach
Letting data drive estimation
No (false) distributional assumptions about returns.
Outlying (dangerous) returns are represented in loss distribution, whereas they might be ignored
with parametric estimation. Problems?
10/13
The Bootstrap
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
11/13
The Bootstrap
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
One problem with NP approach- there is no sampling variation, and no inference possible, nor needed.
11/13
The Bootstrap
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
One problem with NP approach- there is no sampling variation, and no inference possible, nor needed.
Bootstrap gives way of understanding how precise/imprecise the estimates are.
11/13
The Bootstrap
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
One problem with NP approach- there is no sampling variation, and no inference possible, nor needed.
Bootstrap gives way of understanding how precise/imprecise the estimates are.
Approach – the data we observe is just one possible set of data
11/13
The Bootstrap
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
One problem with NP approach- there is no sampling variation, and no inference possible, nor needed.
Bootstrap gives way of understanding how precise/imprecise the estimates are.
Approach – the data we observe is just one possible set of data
Other sets of data possible to acquire – by repeatedly sampling from the data
11/13
The Bootstrap
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
One problem with NP approach- there is no sampling variation, and no inference possible, nor needed.
Bootstrap gives way of understanding how precise/imprecise the estimates are.
Approach – the data we observe is just one possible set of data
Other sets of data possible to acquire – by
repeatedly sampling from the data Two points
11/13
The Bootstrap
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
One problem with NP approach- there is no sampling variation, and no inference possible, nor needed.
Bootstrap gives way of understanding how precise/imprecise the estimates are.
Approach – the data we observe is just one possible set of data
Other sets of data possible to acquire – by repeatedly sampling from the data
Two points
Each sample is the same size as the original data
11/13
The Bootstrap
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
One problem with NP approach- there is no sampling variation, and no inference possible, nor needed.
Bootstrap gives way of understanding how precise/imprecise the estimates are.
Approach – the data we observe is just one possible set of data
Other sets of data possible to acquire – by repeatedly sampling from the data
Two points
Each sample is the same size as the original data Sampling is done with replacement. Thus samples
will differ. Thus Risk Measures will differ.
11/13
The Bootstrap
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
One problem with NP approach- there is no sampling variation, and no inference possible, nor needed.
Bootstrap gives way of understanding how precise/imprecise the estimates are.
Approach – the data we observe is just one possible set of data
Other sets of data possible to acquire – by repeatedly sampling from the data
Two points
Each sample is the same size as the original data Sampling is done with replacement. Thus samples
will differ. Thus Risk Measures will differ.
for each sample, calculate statistic of interest (eg
VaR)
11/13
The Bootstrap
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
One problem with NP approach- there is no sampling variation, and no inference possible, nor needed.
Bootstrap gives way of understanding how precise/imprecise the estimates are.
Approach – the data we observe is just one possible set of data
Other sets of data possible to acquire – by repeatedly sampling from the data
Two points
Each sample is the same size as the original data Sampling is done with replacement. Thus samples
will differ. Thus Risk Measures will differ.
for each sample, calculate statistic of interest (eg
VaR)
This way, can learn about the distribution of the
statistic.
11/13
The Bootstrap: illustration
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
12/13
The Bootstrap: illustration
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
For the single asset, the Var95 estimate from the original population was Var95 = 15.91793.
12/13
The Bootstrap: illustration
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
For the single asset, the Var95 estimate from the original population was Var95 = 15.91793.
Bootstrapping 1000 times gives Bootstrapped VaR95
14 15 16 17 18 19
VaR 95
12/13
Frequency
0 20 40 60 80 100
The Bootstrap: illustration
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
For the single asset, the Var95 estimate from the original population was Var95 = 15.91793.
Bootstrapping 1000 times gives Bootstrapped VaR95
14 15 16 17 18 19
VaR 95
Vertical line is original estimate
12/13
Frequency
0 20 40 60 80 100
The Bootstrap: illustration
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
For the single asset, the Var95 estimate from the original population was Var95 = 15.91793.
Bootstrapping 1000 times gives Bootstrapped VaR95
14 15 16 17 18 19
VaR 95
Vertical line is original estimate
Note variation in VaR with alternative samples.
12/13
Frequency
0 20 40 60 80 100
The Bootstrap – illustration
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
13/13
The Bootstrap – illustration
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
For the two asset portfolio, the Var95 estimate from the original population was
Var95 = 14.22643
13/13
The Bootstrap – illustration
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
For the two asset portfolio, the Var95 estimate from the original population was
Var95 = 14.22643
Bootstrapping 1000 times gives Bootstrapped VaR95
12 13 14 15 16
VaR 95
13/13
Frequency
0 50 100 150
The Bootstrap – illustration
Non- Parametric Methods
Non- Parametric Methods – examples
Non-Parametric Regression Non-Parametric Density Estimation
Historical Simulation
For the two asset portfolio, the Var95 estimate from the original population was
Var95 = 14.22643
Bootstrapping 1000 times gives Bootstrapped VaR95
12 13 14 15 16
VaR 95
Vertical line is original estimate
13/13
Frequency
0 50 100 150