# Copyright (c) 2016 Johan Dahlin [ johan.dahlin (at) liu.se ] # Distributed under the MIT license. # ############################################################################## ############################################################################## import numpy as np from state import smc from para import pmh_correlatedRVs from models import normalIID_2parameters ############################################################################## # Arrange the data structures ############################################################################## sm = smc.smcSampler(); pmh = pmh_correlatedRVs.stcPMH(); ############################################################################## # Setup the system ############################################################################## sys = normalIID_2parameters.ssm() sys.par = np.zeros((sys.nPar,1)) sys.par[0] = 0.50; sys.par[1] = 0.30; sys.par[2] = 0.10; sys.T = 10; sys.xo = 0.0; ##############################################################################
# # (c) 2016 Johan Dahlin # liu (at) johandahlin.com # ############################################################################## ############################################################################## import numpy as np from state import smc from para import gpo_gpy from models import hwsv_4parameters ############################################################################## # Arrange the data structures ############################################################################## sm = smc.smcSampler() gpo = gpo_gpy.stGPO() ############################################################################## # Setup the system ############################################################################## sys = hwsv_4parameters.ssm() sys.par = np.zeros((sys.nPar, 1)) sys.par[0] = 0.20 sys.par[1] = 0.90 sys.par[2] = 0.15 sys.par[3] = -0.70 sys.T = 1000 sys.xo = 0.0 sys.version = "standard"
def estVol(version, data, theta, settings):
# Arrange the data structures
sm = smc.smcSampler()
#=========================================================================
# Setup the system model and load data
#=========================================================================
if (version == 'GSV'):
sys = hwsv_4parameters.ssm()
else:
sys = hwsvalpha_4parameters.ssm()
sys.par = np.zeros((sys.nPar, 1))
sys.xo = 0.0
sys.T = len(data)
sys.version = "standard"
sys.transformY = "none"
# Load data
if (version == 'GSV'):
sys.generateData()
else:
sys.par[3] = 2.0
sys.generateData()
sys.ynoiseless = np.array(data).reshape((sys.T, 1))
sys.y = np.array(data).reshape((sys.T, 1))
#=========================================================================
# Setup the model for inference
#=========================================================================
if (version == 'GSV'):
th = hwsv_4parameters.ssm()
th.nParInference = 3
th.transformY = "none"
else:
th = hwsvalpha_4parameters.ssm()
th.nParInference = 4
th.transformY = "arctan"
th.version = "standard"
th.copyData(sys)
#=========================================================================
# Setup the SMC algorithm
#=========================================================================
if (version == 'GSV'):
sm.filter = sm.bPF
else:
sm.filter = sm.bPFabc
sm.weightdist = settings['smc_weightdist']
sm.tolLevel = settings['smc_tolLevel']
# Add noise to data for noisy ABC
th.makeNoisy(sm)
sm.nPart = settings['smc_nPart']
sm.genInitialState = True
# Estimate the log-volatility
th.storeParameters(theta, sys)
sm.filter(th)
# Return the log-volatility estimate
return sm.xhatf[:, 0]
def estModel(version, data, settings):
# Arrange the data structures
gpo = gpo_gpy.stGPO()
sm = smc.smcSampler()
#=========================================================================
# Setup the system model and load data
#=========================================================================
if (version == 'GSV'):
sys = hwsv_4parameters.ssm()
else:
sys = hwsvalpha_4parameters.ssm()
sys.par = np.zeros((sys.nPar, 1))
sys.xo = 0.0
sys.T = len(data)
sys.version = "standard"
sys.transformY = "none"
# Load data
if (version == 'GSV'):
sys.generateData()
else:
sys.par[3] = 2.0
sys.generateData()
sys.ynoiseless = np.array(data).reshape((sys.T, 1))
sys.y = np.array(data).reshape((sys.T, 1))
#=========================================================================
# Setup the model for inference
#=========================================================================
if (version == 'GSV'):
th = hwsv_4parameters.ssm()
th.transformY = "none"
th.nParInference = 3
else:
th = hwsvalpha_4parameters.ssm()
th.transformY = "arctan"
th.nParInference = 4
th.version = "standard"
th.copyData(sys)
#=========================================================================
# Setup the SMC algorithm
#=========================================================================
if (version == 'GSV'):
sm.filter = sm.bPF
else:
sm.filter = sm.bPFabc
sm.weightdist = settings['smc_weightdist']
sm.tolLevel = settings['smc_tolLevel']
# Add noise to data for noisy ABC
th.makeNoisy(sm)
sm.nPart = settings['smc_nPart']
sm.genInitialState = True
#=========================================================================
# Setup the GPO algorithm
#=========================================================================
gpo.verbose = True
gpo.initPar = settings['gpo_initPar'][0:th.nParInference]
gpo.upperBounds = settings['gpo_upperBounds'][0:th.nParInference]
gpo.lowerBounds = settings['gpo_lowerBounds'][0:th.nParInference]
gpo.preIter = settings['gpo_preIter']
gpo.maxIter = settings['gpo_maxIter']
gpo.jitteringCovariance = 0.01 * np.diag(np.ones(th.nParInference))
gpo.preSamplingMethod = "latinHyperCube"
gpo.EstimateHyperparametersInterval = settings['gpo_estHypParInterval']
gpo.EstimateThHatEveryIteration = False
gpo.EstimateHessianEveryIteration = False
# Estimate parameters
gpo.bayes(sm, sys, th)
gpo.estimateHessian()
# Estimate the log-volatility
th.storeParameters(gpo.thhat, sys)
sm.filter(th)
return sm.xhatf[:, 0], gpo.thhat, np.diag(-gpo.invHessianEstimate)
def estVol(version, data, theta, settings):
# Arrange the data structures
sm = smc.smcSampler()
#=========================================================================
# Setup the system model and load data
#=========================================================================
if (version == 'GSV'):
sys = hwsv_4parameters.ssm()
else:
sys = hwsvalpha_4parameters.ssm()
sys.par = np.zeros((sys.nPar, 1))
sys.xo = 0.0
sys.T = len(data)
sys.version = "standard"
sys.transformY = "none"
# Load data
if (version == 'GSV'):
sys.generateData()
else:
sys.par[3] = 2.0
sys.generateData()
sys.ynoiseless = np.array(data).reshape((sys.T, 1))
sys.y = np.array(data).reshape((sys.T, 1))
#=========================================================================
# Setup the model for inference
#=========================================================================
if (version == 'GSV'):
th = hwsv_4parameters.ssm()
th.nParInference = 3
th.transformY = "none"
else:
th = hwsvalpha_4parameters.ssm()
th.nParInference = 4
th.transformY = "arctan"
th.version = "standard"
th.copyData(sys)
#=========================================================================
# Setup the SMC algorithm
#=========================================================================
if (version == 'GSV'):
sm.filter = sm.bPF
else:
sm.filter = sm.bPFabc
sm.weightdist = settings['smc_weightdist']
sm.tolLevel = settings['smc_tolLevel']
# Add noise to data for noisy ABC
th.makeNoisy(sm)
sm.nPart = settings['smc_nPart']
sm.genInitialState = True
# Estimate the log-volatility
th.storeParameters(theta, sys)
sm.filter(th)
# Return the log-volatility estimate
return sm.xhatf[:, 0]
def estModel(version, data, settings):
# Arrange the data structures
gpo = gpo_gpy.stGPO()
sm = smc.smcSampler()
#=========================================================================
# Setup the system model and load data
#=========================================================================
if (version == 'GSV'):
sys = hwsv_4parameters.ssm()
else:
sys = hwsvalpha_4parameters.ssm()
sys.par = np.zeros((sys.nPar, 1))
sys.xo = 0.0
sys.T = len(data)
sys.version = "standard"
sys.transformY = "none"
# Load data
if (version == 'GSV'):
sys.generateData()
else:
sys.par[3] = 2.0
sys.generateData()
sys.ynoiseless = np.array(data).reshape((sys.T, 1))
sys.y = np.array(data).reshape((sys.T, 1))
#=========================================================================
# Setup the model for inference
#=========================================================================
if (version == 'GSV'):
th = hwsv_4parameters.ssm()
th.transformY = "none"
th.nParInference = 3
else:
th = hwsvalpha_4parameters.ssm()
th.transformY = "arctan"
th.nParInference = 4
th.version = "standard"
th.copyData(sys)
#=========================================================================
# Setup the SMC algorithm
#=========================================================================
if (version == 'GSV'):
sm.filter = sm.bPF
else:
sm.filter = sm.bPFabc
sm.weightdist = settings['smc_weightdist']
sm.tolLevel = settings['smc_tolLevel']
# Add noise to data for noisy ABC
th.makeNoisy(sm)
sm.nPart = settings['smc_nPart']
sm.genInitialState = True
#=========================================================================
# Setup the GPO algorithm
#=========================================================================
gpo.verbose = True
gpo.initPar = settings['gpo_initPar'][0:th.nParInference]
gpo.upperBounds = settings['gpo_upperBounds'][0:th.nParInference]
gpo.lowerBounds = settings['gpo_lowerBounds'][0:th.nParInference]
gpo.preIter = settings['gpo_preIter']
gpo.maxIter = settings['gpo_maxIter']
gpo.jitteringCovariance = 0.01 * np.diag(np.ones(th.nParInference))
gpo.preSamplingMethod = "latinHyperCube"
gpo.EstimateHyperparametersInterval = settings['gpo_estHypParInterval']
gpo.EstimateThHatEveryIteration = False
gpo.EstimateHessianEveryIteration = False
# Estimate parameters
gpo.bayes(sm, sys, th)
gpo.estimateHessian()
# Estimate the log-volatility
th.storeParameters(gpo.thhat, sys)
sm.filter(th)
return sm.xhatf[:, 0], gpo.thhat, np.diag(-gpo.invHessianEstimate)
