nonlinear models for volatility prediction in the financial markets
DESCRIPTION
TRANSCRIPT
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Nonlinear Modelsfor Volatility Predictionin the Financial Markets
Matteo Ainardi
Advisors: Prof. Gianpiero Cabodi, Prof. Derong Liu
University of Illinois at ChicagoMaster of Science in Electrical and Computer Engineering
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Outline
1 IntroductionVolatility
2 Prediction ModelsVolatility Models
3 ResultsVolatility PredictionInvestment Strategy
4 Modeling GUI
5 Conclusions
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility
Volatility
Thesis goal
Development and evaluation of models for volatility prediction.
Qualitative definition
Volatility: degree of price variation over time.
Why Volatility?
Investors assess expected returns of an asset against its risk.
Financial institutions want to ensure that the value of theirassets does not fall below some minimum level that wouldexpose them to insolvency.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility
Volatility
Thesis goal
Development and evaluation of models for volatility prediction.
Qualitative definition
Volatility: degree of price variation over time.
Why Volatility?
Investors assess expected returns of an asset against its risk.
Financial institutions want to ensure that the value of theirassets does not fall below some minimum level that wouldexpose them to insolvency.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility
Volatility
Thesis goal
Development and evaluation of models for volatility prediction.
Qualitative definition
Volatility: degree of price variation over time.
Why Volatility?
Investors assess expected returns of an asset against its risk.
Financial institutions want to ensure that the value of theirassets does not fall below some minimum level that wouldexpose them to insolvency.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility
Volatility
Thesis goal
Development and evaluation of models for volatility prediction.
Qualitative definition
Volatility: degree of price variation over time.
Why Volatility?
Investors assess expected returns of an asset against its risk.
Financial institutions want to ensure that the value of theirassets does not fall below some minimum level that wouldexpose them to insolvency.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility
Volatility
Volatility Features
not directly observable
not constant over time
Quantitative Definition
Time varying volatility measure from the price time series:
σ2t = (rt − r)2
rt= pt−pt−1
pt−1: return on day t,
r : mean return over the last 200 days period.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility
Volatility
Volatility Features
not directly observable
not constant over time
Quantitative Definition
Time varying volatility measure from the price time series:
σ2t = (rt − r)2
rt= pt−pt−1
pt−1: return on day t,
r : mean return over the last 200 days period.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility
Volatility
Volatility Features
not directly observable
not constant over time
Quantitative Definition
Time varying volatility measure from the price time series:
σ2t = (rt − r)2
rt= pt−pt−1
pt−1: return on day t,
r : mean return over the last 200 days period.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility
Volatility
Volatility Features
not directly observable
not constant over time
Quantitative Definition
Time varying volatility measure from the price time series:
σ2t = (rt − r)2
rt= pt−pt−1
pt−1: return on day t,
r : mean return over the last 200 days period.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Prediction Models
Let’s assume that the system to forecast can be described by aregression equation of the form
yt+1 = f0(ϕt) + dt
ϕt = [rt , . . . , rt−N+1, yt , . . . , yt−M+1]
t ∈ N: time [days],
yt : volatility at day t,
ϕt : regressor,
dt : noise affecting the system.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Prediction Models
Let’s assume that the system to forecast can be described by aregression equation of the form
yt+1 = f0(ϕt) + dt
ϕt = [rt , . . . , rt−N+1, yt , . . . , yt−M+1]
t ∈ N: time [days],
yt : volatility at day t,
ϕt : regressor,
dt : noise affecting the system.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Prediction Models
Let’s assume that the system to forecast can be described by aregression equation of the form
yt+1 = f0(ϕt) + dt
ϕt = [rt , . . . , rt−N+1, yt , . . . , yt−M+1]
t ∈ N: time [days],
yt : volatility at day t,
ϕt : regressor,
dt : noise affecting the system.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Prediction Models
Let’s assume that the system to forecast can be described by aregression equation of the form
yt+1 = f0(ϕt) + dt
ϕt = [rt , . . . , rt−N+1, yt , . . . , yt−M+1]
t ∈ N: time [days],
yt : volatility at day t,
ϕt : regressor,
dt : noise affecting the system.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Prediction Models
Let’s assume that the system to forecast can be described by aregression equation of the form
yt+1 = f0(ϕt) + dt
ϕt = [rt , . . . , rt−N+1, yt , . . . , yt−M+1]
t ∈ N: time [days],
yt : volatility at day t,
ϕt : regressor,
dt : noise affecting the system.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Prediction Models
A prediction model f can be defined as an approximation of f0,providing a prediction yt+1 of yt+1:
yt+1 = f (ϕt),
where ϕt represents an estimate of the true regressor ϕt .
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Prediction Models - Statistical/Parametric Approach
The traditional approach followed to build a model implies a choiceof a specific structure for the functional form f0 and statisticalassumptions on the noise dt affecting the system.
if possible, physical/economical laws are used to obtain aparametric representation of the system f (ϕ, θ)
as a parametric combination of basis functions (polynomial,sigmoid, ...)
The parameters θ are then estimated from data by optimizingLeast Squares or Maximum Likelihood functions.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
GARCH Models
Linear Regression / Moving Average Models
Neural Network Models
Nonlinear Set Membership Models
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
GARCH Models
Generalized Autoregressive Conditional HeteroskedasticityThis methodology is the most widely adopted and led its creator,Prof. Robert Engle, to the Nobel Prize in Economy in 2003.
The volatility prediction is based on
long run constantvolatility
most recent return
previous volatilityprediction
yt+1 = ω + αr2t + βyt
The GARCH regressor is ϕt = [rt , yt ]
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
GARCH Models
Generalized Autoregressive Conditional HeteroskedasticityThis methodology is the most widely adopted and led its creator,Prof. Robert Engle, to the Nobel Prize in Economy in 2003.The volatility prediction is based on
long run constantvolatility
most recent return
previous volatilityprediction
yt+1 = ω + αr2t + βyt
The GARCH regressor is ϕt = [rt , yt ]
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
GARCH Models
Generalized Autoregressive Conditional HeteroskedasticityThis methodology is the most widely adopted and led its creator,Prof. Robert Engle, to the Nobel Prize in Economy in 2003.The volatility prediction is based on
long run constantvolatility
most recent return
previous volatilityprediction
yt+1 = ω + αr2t + βyt
The GARCH regressor is ϕt = [rt , yt ]
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
GARCH Models
Generalized Autoregressive Conditional HeteroskedasticityThis methodology is the most widely adopted and led its creator,Prof. Robert Engle, to the Nobel Prize in Economy in 2003.The volatility prediction is based on
long run constantvolatility
most recent return
previous volatilityprediction
yt+1 = ω + αr2t + βyt
The GARCH regressor is ϕt = [rt , yt ]
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
GARCH Models
Generalized Autoregressive Conditional HeteroskedasticityThis methodology is the most widely adopted and led its creator,Prof. Robert Engle, to the Nobel Prize in Economy in 2003.The volatility prediction is based on
long run constantvolatility
most recent return
previous volatilityprediction
yt+1 = ω + αr2t + βyt
The GARCH regressor is ϕt = [rt , yt ]
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models - Linear Models
Linear Regression Models
yt+1 =N−1∑i=0
αi r2t−i +
M−1∑j=0
βj yt−j
ϕt = [rt , . . . , rt−N+1, yt , . . . , yt−M+1]
αi and βi are the parameters that must be estimated.
Moving Average Models
yt+1 =1
N
N−1∑i=0
r2t−i , ϕt = [rt , . . . , rt−N+1]
Particular case of linear regression: αi = 1N , βj = 0 for each i , j .
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models - Linear Models
Linear Regression Models
yt+1 =N−1∑i=0
αi r2t−i +
M−1∑j=0
βj yt−j
ϕt = [rt , . . . , rt−N+1, yt , . . . , yt−M+1]
αi and βi are the parameters that must be estimated.
Moving Average Models
yt+1 =1
N
N−1∑i=0
r2t−i , ϕt = [rt , . . . , rt−N+1]
Particular case of linear regression: αi = 1N , βj = 0 for each i , j .
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models - Neural Networks
Neural Network Models
yt+1 =r∑
i=1
αiσ(βiϕt − λi ) + ζ
r : number of neurons in the network
σ(x) = 21+e−2x : sigmoidal function (neuron transfer function)
ϕt = [rt , . . . , rt−N+1, yt , . . . , yt−M+1]
αi , βi , λi : parameters estimated through a network trainingalgorithm
ζ: noise affecting the system
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
Observations
These models require the selection of a specific functional form off0 and statistical assumptions on the noise (usually supposed to begaussian).
Search of the appropriate functional form and parameters:complex and computationally heavy
Wrong f0 choice ⇒ degradation in model accuracy
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
Observations
These models require the selection of a specific functional form off0 and statistical assumptions on the noise (usually supposed to begaussian).
Search of the appropriate functional form and parameters:complex and computationally heavy
Wrong f0 choice ⇒ degradation in model accuracy
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
Observations
These models require the selection of a specific functional form off0 and statistical assumptions on the noise (usually supposed to begaussian).
Search of the appropriate functional form and parameters:complex and computationally heavy
Wrong f0 choice ⇒ degradation in model accuracy
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models - Nonlinear Set Membership
NSM Methodology
Nonlinear Set Membership models do not require the choice of aparametric form of f0, but more relaxed assumptions on theanalyzed system:
bound on its derivatives
bound on the noise affecting the system
Advantages:
avoid the complexity/accuracy problems posed by the choiceof the model function and of the parameters.
NSM Local Approach: check if a given model can be improvedby identifying a NSM model to forecast its residual process.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models - Nonlinear Set Membership
NSM Methodology
Nonlinear Set Membership models do not require the choice of aparametric form of f0, but more relaxed assumptions on theanalyzed system:
bound on its derivatives
bound on the noise affecting the system
Advantages:
avoid the complexity/accuracy problems posed by the choiceof the model function and of the parameters.
NSM Local Approach: check if a given model can be improvedby identifying a NSM model to forecast its residual process.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models - Nonlinear Set Membership
NSM Methodology
Nonlinear Set Membership models do not require the choice of aparametric form of f0, but more relaxed assumptions on theanalyzed system:
bound on its derivatives
bound on the noise affecting the system
Advantages:
avoid the complexity/accuracy problems posed by the choiceof the model function and of the parameters.
NSM Local Approach: check if a given model can be improvedby identifying a NSM model to forecast its residual process.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models - Nonlinear Set Membership
NSM Models
yt+1 =1
2(f (ϕt) + f (ϕt)),
f (ϕT ) = mink=1,...,T−1
(yk+1 + εk + γ ‖ϕT − ϕk‖)
f (ϕT ) = maxk=1,...,T−1
(yk+1 − εk − γ ‖ϕT − ϕk‖)
where γ (gradient) and ε (noise) bounds are derived from pastmeasured data through a validation procedure.NSM Regressor: ϕt = [rt , . . . , rt−N+1, yt , . . . , yt−M+1]
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models - Nonlinear Set Membership
NSM Models
yt+1 =1
2(f (ϕt) + f (ϕt)),
f (ϕT ) = mink=1,...,T−1
(yk+1 + εk + γ ‖ϕT − ϕk‖)
f (ϕT ) = maxk=1,...,T−1
(yk+1 − εk − γ ‖ϕT − ϕk‖)
where γ (gradient) and ε (noise) bounds are derived from pastmeasured data through a validation procedure.NSM Regressor: ϕt = [rt , . . . , rt−N+1, yt , . . . , yt−M+1]
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models - Nonlinear Set Membership
NSM Models
yt+1 =1
2(f (ϕt) + f (ϕt)),
f (ϕT ) = mink=1,...,T−1
(yk+1 + εk + γ ‖ϕT − ϕk‖)
f (ϕT ) = maxk=1,...,T−1
(yk+1 − εk − γ ‖ϕT − ϕk‖)
where γ (gradient) and ε (noise) bounds are derived from pastmeasured data through a validation procedure.NSM Regressor: ϕt = [rt , . . . , rt−N+1, yt , . . . , yt−M+1]
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models - Nonlinear Set Membership
NSM Models
yt+1 =1
2(f (ϕt) + f (ϕt)),
f (ϕT ) = mink=1,...,T−1
(yk+1 + εk + γ ‖ϕT − ϕk‖)
f (ϕT ) = maxk=1,...,T−1
(yk+1 − εk − γ ‖ϕT − ϕk‖)
where γ (gradient) and ε (noise) bounds are derived from pastmeasured data through a validation procedure.
NSM Regressor: ϕt = [rt , . . . , rt−N+1, yt , . . . , yt−M+1]
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models - Nonlinear Set Membership
NSM Models
yt+1 =1
2(f (ϕt) + f (ϕt)),
f (ϕT ) = mink=1,...,T−1
(yk+1 + εk + γ ‖ϕT − ϕk‖)
f (ϕT ) = maxk=1,...,T−1
(yk+1 − εk − γ ‖ϕT − ϕk‖)
where γ (gradient) and ε (noise) bounds are derived from pastmeasured data through a validation procedure.NSM Regressor: ϕt = [rt , . . . , rt−N+1, yt , . . . , yt−M+1]
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models - Nonlinear Set Membership
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
Example
Availablemeasurements:
t ϕt yt+1
1 [3, 3] 3
2 [1, 3] 4
3 [3, 2] 5
Let’s assume
γ = 0.5, εr = 0.2
ϕ4 = [2, 1].
How to determine thevalue of y5?
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
Example
1) compute εt
t yt εt
2 3 0.6
3 4 0.8
4 5 1.0
2) compute ‖ϕ4 − ϕt‖
t ϕt ‖ϕ4 − ϕt‖1 [3, 3] 2.24
2 [1, 3] 2.24
3 [3, 2] 1.41
f (ϕ4) = mink=1,...,3
(yk+1 + εk + γ ‖ϕ4 − ϕk‖)
f (ϕ4) = maxk=1,...,3
(yk+1 − εk − γ ‖ϕ4 − ϕk‖)
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
Example
1) compute εt
t yt εt
2 3 0.6
3 4 0.8
4 5 1.0
2) compute ‖ϕ4 − ϕt‖
t ϕt ‖ϕ4 − ϕt‖1 [3, 3] 2.24
2 [1, 3] 2.24
3 [3, 2] 1.41
f (ϕ4) = mink=1,...,3
(yk+1 + εk + γ ‖ϕ4 − ϕk‖)
f (ϕ4) = maxk=1,...,3
(yk+1 − εk − γ ‖ϕ4 − ϕk‖)
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
Example
1) compute εt
t yt εt
2 3 0.6
3 4 0.8
4 5 1.0
2) compute ‖ϕ4 − ϕt‖
t ϕt ‖ϕ4 − ϕt‖1 [3, 3] 2.24
2 [1, 3] 2.24
3 [3, 2] 1.41
f (ϕ4) = mink=1,...,3
(yk+1 + εk + γ ‖ϕ4 − ϕk‖)
f (ϕ4) = maxk=1,...,3
(yk+1 − εk − γ ‖ϕ4 − ϕk‖)
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
Example
f (ϕ4) = min (4.72, 5.92, 6.71) = 4.72
f (ϕ4) = max (1.28, 2.08, 3.30) = 3.30
The NSM prediction for y5 value will be
y5 =1
2(f (ϕ4) + f (ϕ4)) = 4.01.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models
Example
f (ϕ4) = min (4.72, 5.92, 6.71) = 4.72
f (ϕ4) = max (1.28, 2.08, 3.30) = 3.30
The NSM prediction for y5 value will be
y5 =1
2(f (ϕ4) + f (ϕ4)) = 4.01.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility Models
Volatility Models - Nonlinear Set Membership
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility PredictionInvestment Strategy
Models Results
Models are evaluated on real financial time series:
FDAX
IBM
both in terms of predictive performance and by simulating aninvestment strategy.
IBM Dataset
The IBM dataset is composed of 2248 daily closure prices, from2000 to 2009.
Identification dataset: samples 1 to 1500.
Validation dataset: samples 1501 to 2248.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility PredictionInvestment Strategy
Results Evaluation
Trivial Models
Persistent Modelyt+1 = yt ,
Constant Modelyt+1 = const,
Exact Modelyt+1 = yt+1
These models are used as terms of comparisons.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility PredictionInvestment Strategy
IBM Volatility
Performance Evaluation
Model RMSE R2MZ
Persistent 10.66× 10−4 2.84%
GARCH 7.82× 10−4 12.20%
MAV 7.74× 10−4 13.59%
NN 7.68× 10−4 14.03%
NSM 7.86× 10−4 12.77%
RMSE (Root Mean Squared Error)
RMSE =
√∑nt=1(yt − yt)2
n
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility PredictionInvestment Strategy
IBM Volatility
Performance Evaluation
Model RMSE R2MZ
Persistent 10.66× 10−4 2.84%
GARCH 7.82× 10−4 12.20%
MAV 7.74× 10−4 13.59%
NN 7.68× 10−4 14.03%
NSM 7.86× 10−4 12.77%
RMSE (Root Mean Squared Error)
RMSE =
√∑nt=1(yt − yt)2
n
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility PredictionInvestment Strategy
IBM Volatility
Performance Evaluation
Model RMSE R2MZ
Persistent 10.66× 10−4 2.84%
GARCH 7.82× 10−4 12.20%
MAV 7.74× 10−4 13.59%
NN 7.68× 10−4 14.03%
NSM 7.86× 10−4 12.77%
R2MZ (Mincer Zarnowitz’s Regression R2)It is widely adopted in econometrics to assessthe practical utility of the volatility model.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility PredictionInvestment Strategy
Investment Strategy
Scenario
An investor can choose among a risk free asset which gives aconstant return and a risky asset.The composition of the investor’s portfolio can be changed everyday.
Investment Strategy
The investor accepts a certain level of volatility for its portfolio ifthere is the opportunity of an higher profit.
↘ predicted volatility ⇒↗ risky asset weight
↗ predicted volatility ⇒↘ risky asset weight
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility PredictionInvestment Strategy
Investment Strategy
Scenario
An investor can choose among a risk free asset which gives aconstant return and a risky asset.The composition of the investor’s portfolio can be changed everyday.
Investment Strategy
The investor accepts a certain level of volatility for its portfolio ifthere is the opportunity of an higher profit.
↘ predicted volatility ⇒↗ risky asset weight
↗ predicted volatility ⇒↘ risky asset weight
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility PredictionInvestment Strategy
Investment Strategy Results
Strategy Profit (×10−3) ∆(×10−3) Risk (×10−3) ProfitRisk
Constant 65.7 — 142 0.47
Exact Vol 163 97.3 141 1.16
Persistent 54.4 −11.3 348 0.16
GARCH 110 44.3 158 0.70
MAV 126 60.3 179 0.70
NN 115 49.3 162 0.71
NSM 152 86.3 216 0.70
Profit: mean portfolio return
∆: improvement with respect to the constant strategy
Risk: portfolio return standard deviation
Profit refers to the annualized mean portfolio return.For example the NSM strategy showsabout a +15% annual portfolio variation.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility PredictionInvestment Strategy
Investment Strategy Results
Strategy Profit (×10−3) ∆(×10−3) Risk (×10−3) ProfitRisk
Constant 65.7 — 142 0.47
Exact Vol 163 97.3 141 1.16
Persistent 54.4 −11.3 348 0.16
GARCH 110 44.3 158 0.70
MAV 126 60.3 179 0.70
NN 115 49.3 162 0.71
NSM 152 86.3 216 0.70
Profit: mean portfolio return
∆: improvement with respect to the constant strategy
Risk: portfolio return standard deviation
Profit refers to the annualized mean portfolio return.For example the NSM strategy showsabout a +15% annual portfolio variation.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility PredictionInvestment Strategy
Investment Strategy Results
Strategy Profit (×10−3) ∆(×10−3) Risk (×10−3) ProfitRisk
Constant 65.7 — 142 0.47
Exact Vol 163 97.3 141 1.16
Persistent 54.4 −11.3 348 0.16
GARCH 110 44.3 158 0.70
MAV 126 60.3 179 0.70
NN 115 49.3 162 0.71
NSM 152 86.3 216 0.70
Profit: mean portfolio return
∆: improvement with respect to the constant strategy
Risk: portfolio return standard deviation
Profit refers to the annualized mean portfolio return.For example the NSM strategy showsabout a +15% annual portfolio variation.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Volatility PredictionInvestment Strategy
Investment Strategy Results
Strategy Profit (×10−3) ∆(×10−3) Risk (×10−3) ProfitRisk
Constant 65.7 — 142 0.47
Exact Vol 163 97.3 141 1.16
Persistent 54.4 −11.3 348 0.16
GARCH 110 44.3 158 0.70
MAV 126 60.3 179 0.70
NN 115 49.3 162 0.71
NSM 152 86.3 216 0.70
Profit: mean portfolio return
∆: improvement with respect to the constant strategy
Risk: portfolio return standard deviation
Profit refers to the annualized mean portfolio return.For example the NSM strategy showsabout a +15% annual portfolio variation.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Data Load GUI
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Model Identification GUI
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Automatic Model Identification GUI
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Display Results GUI
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Conclusions
GARCH, MAV, NSM and NN models show significantimprovements with respect to the persistent model and toconstant volatility.
NSM optimality analysis, based on local approach, indicatesthat GARCH models performance can not be significantlyimproved.
NN models often produce unstable results in presence ofvolatility dynamics very different from the ones on which theyare trained.
NSM models are able to produce robust results even inpresence of extreme volatility variations.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Conclusions
GARCH, MAV, NSM and NN models show significantimprovements with respect to the persistent model and toconstant volatility.
NSM optimality analysis, based on local approach, indicatesthat GARCH models performance can not be significantlyimproved.
NN models often produce unstable results in presence ofvolatility dynamics very different from the ones on which theyare trained.
NSM models are able to produce robust results even inpresence of extreme volatility variations.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Conclusions
GARCH, MAV, NSM and NN models show significantimprovements with respect to the persistent model and toconstant volatility.
NSM optimality analysis, based on local approach, indicatesthat GARCH models performance can not be significantlyimproved.
NN models often produce unstable results in presence ofvolatility dynamics very different from the ones on which theyare trained.
NSM models are able to produce robust results even inpresence of extreme volatility variations.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Conclusions
GARCH, MAV, NSM and NN models show significantimprovements with respect to the persistent model and toconstant volatility.
NSM optimality analysis, based on local approach, indicatesthat GARCH models performance can not be significantlyimproved.
NN models often produce unstable results in presence ofvolatility dynamics very different from the ones on which theyare trained.
NSM models are able to produce robust results even inpresence of extreme volatility variations.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Conclusions
All the experiments have been implemented and executed inMatlab, the development of the GUI resulted into about 2500lines of Matlab code.
More than 30 prediction models have been identified andevaluated on the different datasets, relying on differentvolatility measures proposed in financial literature.
Future developments: try different regressor forms, apply themodels to a larger number of markets and timeframes.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Conclusions
All the experiments have been implemented and executed inMatlab, the development of the GUI resulted into about 2500lines of Matlab code.
More than 30 prediction models have been identified andevaluated on the different datasets, relying on differentvolatility measures proposed in financial literature.
Future developments: try different regressor forms, apply themodels to a larger number of markets and timeframes.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Conclusions
All the experiments have been implemented and executed inMatlab, the development of the GUI resulted into about 2500lines of Matlab code.
More than 30 prediction models have been identified andevaluated on the different datasets, relying on differentvolatility measures proposed in financial literature.
Future developments: try different regressor forms, apply themodels to a larger number of markets and timeframes.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Conclusions
All the experiments have been implemented and executed inMatlab, the development of the GUI resulted into about 2500lines of Matlab code.
More than 30 prediction models have been identified andevaluated on the different datasets, relying on differentvolatility measures proposed in financial literature.
Future developments: try different regressor forms, apply themodels to a larger number of markets and timeframes.
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets
OutlineIntroduction
Prediction ModelsResults
Modeling GUIConclusions
Thank You
Matteo Ainardi Nonlinear Models for Volatility Prediction in the Financial Markets