realized volatility estimation (slides)
DESCRIPTION
Barcelona GSE Master Project by Miquel Masoliver, Guillem Roig, Shikhar Singla Master Program: Finance About Barcelona GSE master programs: http://j.mp/MastersBarcelonaGSETRANSCRIPT
Realized Volatility Estimation
Miquel Masoliver Guillem Roig Shikhar Singla
July 1, 2014
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 1 / 16
Roadmap
1 Introduction
2 Realized Volatility Estimator
3 Two-Scaled Realized Volatility Estimator
4 Realized Kernel
5 Threshold Realized Variance
6 Simulation setup
7 Data implementation
8 Conclusions
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 2 / 16
Introduction
Volatility is the main driver of portfolio construction, option pricing anddetermination of a firm’s exposure to risk
Parametric models such as the ARCH family rely upon specific distributionalassumptions, immediately calling to question its robustness
Use a non-parametric approach based on summing squares and cross-productsof intraday high-frequency returns to construct estimates of realized dailyvolatility, quadratic variation is used as an ex-post variation of asset prices
The main purpose of this study is to try to find the optimal volatilityestimator in a non-parametric framework
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 3 / 16
Realized Volatility Estimator
Estimates the quadratic variation of prices over some time interval
Discretize the time interval τ = [0, 1] into a grid of subintervalst0 = 0 ≤ · · · ≤ tn = 1 of length ti − ti−1 = ∆ = 1/n, and then set the pricesyi = yti .
RV Estimator
RV 2 =n−1∑i=1
(yi+1 − yi )2
Inconsistent and biased under market microstructure noise and price jumps
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 4 / 16
Two-Scaled Realized Volatility Estimator
Decomposes the total observed variance into two components: fundamentalprice and market microstructure noise
Sample sparsely at some lower frequency to evaluate the quadratic variationat the two frequencies
Consistent and unbiased under market microstructure noise
Two-Scaled Realized Volatility Estimator
TSRV 2 =1
K
K∑k=1
∑ti ,ti+1∈Tk
(yti+1 − yti )2 − n
n
∑tj ,tj+1∈T
(ytj+1 − ytj )2
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 5 / 16
Sample is partitioned into K non-overlapping grids with equal number ofobservations
y1 y2 . . . y50
y51 y52 . . . y100
y101 y102 . . . y150
......
......
y951 y952 . . . y1000
Table : Sampling process
Two-Scaled Realized Volatility Estimator
TSRV 2 =1
K
K∑k=1
∑ti ,ti+1∈Tk
(yti+1 − yti )2 − n
n
∑tj ,tj+1∈T
(ytj+1 − ytj )2
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 6 / 16
Realized Kernel
Linear function of autocovariances weighted by a specific function (Parzen)
Guarantees consistency and positive semi-definiteness
Realized Kernel Estimator
K (X ) =H∑
h=−H
k(h
H + 1)γh,
where γh is the matrix of autocovariances and and we add 2H such matrices
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 7 / 16
Threshold Realized Variance
All previous estimators are inconsistent under jumps
Detect the jumps i.e. returns above a certain threshold, and not includethose price points in volatility calculation where there is a jump
Threshold Realized Variance
TRVδ =
[T/δ]∑j=1
(yj − yj−1)21{(yj−yj−1)2≤Θ(δ)}
TRV1 =
[T/δ]∑j=1
(yj − yj−1)21{(yj−yj−1)2≤log( 1δ )√δ}
TRV2 =
[T/δ]∑j=1
(yj − yj−1)21{(yj−yj−1)2≤√δ}
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 8 / 16
Simulation setup
Prices governed by a Geometric Brownian Motion with Heston Structure forVolatility (true volatility is observable!)
Jumps generated using a compound Poisson process
Market microstructure noise produced by random draws from a normaldistribution with different variances
The number of subsamples for TSRV and number of autocovariance matricesfor RK are chosen so as to minimize the noise effect
Frobenius and Infinite norms to compare between the estimators
Frobenius and Inifinite Norm
Frobenius Norm ||A||f =
√√√√ m∑i=1
m∑j=1
|aij |2
Infinite Norm ||A||i = max|aij |.
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 9 / 16
Figure : Frobenius and Infinite norm to test the performance of the estimators on thebenchmark scenario i.e. neither jumps nor noise
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 10 / 16
Figure : Frobenius and Infinite norms to test the performance of the estimators in thepresence of jumps
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 11 / 16
Figure : Comparison of three estimators using Frobenius and Infinite norms underpresence of noise
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 12 / 16
Data implentation
Tick-by-tick prices of S&P-100 stocks executed on October 27th, 2010
Trading day of 6.5 hours is normalized into a [0,1] interval
To make the price points simultaneous for all stocks, the last recorded tradeis kept
Stocks with daily variance lower than 0.001 are excluded from the dataset
Cannot use norms to assess performance (no true volatility). Construct globalMVP portfolio and compute returns over the next month
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 13 / 16
Figure : Performance of RV, Kernel and uniformly weighted portfolio using MVP
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 14 / 16
Figure : Performance of RV and Threshold RV using MVP
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 15 / 16
Conclusions
When noise is present TSRV and Realized Kernel outperform RealizedVariance
Using the highest available frequency does not necessarily result in the bestestimates
Minimum values for the ex-post portfolios variance are found for samplingfrequencies equal to 30 seconds for RV, TRV1 and TRV2 estimators
Although TRV1 outperforms RV we cannot conclude that jumps do exist inintraday data, since both specifications behave very similarly
High-frequency data achieves better estimates than daily data
Miquel Masoliver, Guillem Roig, Shikhar Singla Realized Volatility Estimation July 1, 2014 16 / 16