def fit_t_fp(x, p=None):
"""For details, see here.
Parameters
----------
x : array, shape (t_,)
p : array, shape (t_,), optional
Returns
-------
nu : float
mu : float
sigma2 : float
"""
t_ = x.shape[0]
if p is None:
p = np.ones(t_) / t_
# Step 1: Compute negative log-likelihood function
def llh(params):
nu, mu, sigma = params
return -p @ t.logpdf(x, nu, mu, sigma)
# Step 2: Find the optimal dof
m, s2 = meancov_sp(x, p)
param0 = [10, m, np.sqrt(s2)]
bnds = [(1e-20, None), (None, None), (1e-20, None)]
nu, mu, sigma = minimize(llh, param0, bounds=bnds)['x']
sigma2 = sigma**2
return nu, mu, sigma2
def ewm_meancov(x, tau_hl, t=None, w=None):
"""For details, see here.
Parameters
----------
x : array, shape (t_, n_) if n_>1 or (t_,) for n=1
tau_hl: scalar
t : int
w : int
Returns
-------
ewma_t_x : array, shape (n_,)
ewm_cv_t_x : array, shape (n_, n_)
"""
t_ = x.shape[0]
x = x.reshape(t_, -1)
if t is None:
t = t_
if w is None:
w = t_
assert (t >= w), "t should be greater or equal to w."
p_w = np.exp(-np.log(2) / tau_hl * np.arange(0, w))[::-1].reshape(-1)
gamma_w = np.sum(p_w)
ewma_t_x, ewm_cv_t_x = meancov_sp(x[t - w:t, :], p_w / gamma_w)
return np.squeeze(ewma_t_x), np.squeeze(ewm_cv_t_x)
def fit_factor_analysis(x, k_, p=None, method='PrincAxFact'):
"""For details, see here.
Parameters
----------
x : array, shape (t_, n_) if n_>1 or (t_, ) for n_=1
k_ : scalar
p : array, shape (t_,), optional
method : string, optional
Returns
-------
alpha_hat : array, shape (n_,)
beta_hat : array, shape (n_, k_) if k_>1 or (n_, ) for k_=1
delta2 : array, shape(n_, n_)
z_reg : array, shape(t_, n_) if n_>1 or (t_, ) for n_=1
"""
t_ = x.shape[0]
if len(x.shape) == 1:
x = x.reshape((t_, 1))
if p is None:
p = np.ones(t_) / t_
# Step 1: Compute HFP mean and covariance of X
m_x_hat_hfp, s2_x_hat_hfp = meancov_sp(x, p)
# Step 2: Estimate alpha
alpha_hat = m_x_hat_hfp
# Step 3: Decompose covariance matrix
if method == 'PrincAxFact' or method.lower() == 'paf':
beta_hat, delta2_hat = factor_analysis_paf(s2_x_hat_hfp, k_)
else:
beta_hat, delta2_hat = factor_analysis_mlf(s2_x_hat_hfp, k_)
if k_ == 1:
beta_hat = beta_hat.reshape(-1, 1)
# Step 4: Compute factor analysis covariance matrix
s2_x_hat_fa = beta_hat @ beta_hat.T + np.diagflat(delta2_hat)
# Step 5: Approximate hidden factor via regression
if np.all(delta2_hat != 0):
omega2 = np.diag(1 / delta2_hat)
z_reg = beta_hat.T @ \
(omega2-omega2@beta_hat@
np.linalg.inv(beta_hat.T@omega2@beta_hat + np.eye(k_))@
beta_hat.T@omega2)@(x-m_x_hat_hfp).T
else:
z_reg = beta_hat.T @ np.linalg.inv(s2_x_hat_fa) @ (x - m_x_hat_hfp).T
return alpha_hat, np.squeeze(beta_hat), delta2_hat, np.squeeze(z_reg.T)
def fit_lfm_lasso(x, z, p=None, lam=1e-2, fit_intercept=True):
"""For details, see here.
Parameters
----------
x : array, shape (t_, n_)
z : array, shape (t_, k_)
p : array, optional, shape (t_,)
lam : float, optional
fit_intercept : bool, optional
Returns
-------
alpha : array, shape (n_,)
beta : array, shape (n_, k_)
s2_u : array, shape (n_, n_)
u : array, shape (t_, n_)
"""
if len(x.shape) == 1:
x = x.reshape(-1, 1)
if len(z.shape) == 1:
z = z.reshape(-1, 1)
t_, n_ = x.shape
k_ = z.shape[1]
if p is None:
p = np.ones(t_) / t_
if lam == 0:
alpha, beta, s2_u, u = fit_lfm_ols(x, z, p, fit_intercept)
else:
if fit_intercept is True:
m_x = p @ x
m_z = p @ z
else:
m_x = np.zeros(n_, )
m_z = np.zeros(k_, )
x_p = ((x - m_x).T * np.sqrt(p)).T
z_p = ((z - m_z).T * np.sqrt(p)).T
clf = Lasso(alpha=lam / (2. * t_), fit_intercept=False)
clf.fit(z_p, x_p)
beta = clf.coef_
if k_ == 1:
alpha = m_x - beta * m_z
u = x - alpha - z * beta
else:
alpha = m_x - beta @ m_z
u = x - alpha - z @ np.atleast_2d(beta).T
_, s2_u = meancov_sp(u, p)
return alpha, beta, s2_u, u
def twist_scenarios_mom_match(x, m_, s2_, p=None, method='Riccati', d=None):
"""For details, see here.
Parameters
----------
x : array, shape (j_,n_) if n_>1 or (j_,) for n_=1
m_ : array, shape (n_,)
s2_ : array, shape (n_,n_)
p : array, optional, shape (j_,)
method : string, optional
d : array, shape (k_, n_), optional
Returns
-------
x : array, shape (j_, n_) if n_>1 or (j_,) for n_=1
"""
if np.ndim(m_) == 0:
m_ = np.reshape(m_, 1).copy()
else:
m_ = np.array(m_).copy()
if np.ndim(s2_) == 0:
s2_ = np.reshape(s2_, (1, 1))
else:
s2_ = np.array(s2_).copy()
if len(x.shape) == 1:
x = x.reshape(-1, 1).copy()
if p is None:
j_ = x.shape[0]
p = np.ones(j_) / j_ # uniform probabilities as default value
# Step 1. Original moments
m_x, s2_x = meancov_sp(x, p)
# Step 2. Transpose-square-root of s2_x
r_x = transpose_square_root(s2_x, method, d)
# Step 3. Transpose-square-root of s2_
r_ = transpose_square_root(s2_, method, d)
# Step 4. Twist matrix
b = r_ @ np.linalg.inv(r_x)
# Step 5. Shift vector
a = m_.reshape(-1, 1) - b @ m_x.reshape(-1, 1)
# Step 6. Twisted scenarios
x_ = (a + b @ x.T).T
return np.squeeze(x_)
def fit_lfm_pcfp(x, p, sig2, k_):
"""For details, see here.
Parameters
----------
x : array, shape (t_, n_)
p : array, shape (t_,)
sig2 : array, shape (t_, t_)
k_ : scalar
Returns
-------
alpha_PC : array, shape (n_,)
beta_PC : array, shape (n_, k_)
gamma_PC : array, shape(n_, k_)
s2_PC : array, shape(n_, n_)
"""
t_, n_ = x.shape
# Step 1: Compute HFP-expectation and covariance of x
m_x, s2_x = meancov_sp(x, p)
# Step 2: Compute the Choleski root of sig2
sig = np.linalg.cholesky(sig2)
# Step 3: Perform spectral decomposition
s2_tmp = np.linalg.solve(sig, (s2_x.dot(np.linalg.pinv(sig))))
e, lambda2 = pca_cov(s2_tmp)
# Step 4: Compute optimal loadings for PC LFM
beta_PC = sig @ e[:, :k_]
# Step 5: Compute factor extraction matrix for PC LFM
gamma_PC = (np.linalg.solve(sig, np.eye(n_))) @ e[:, :k_]
# Step 6: Compute shifting term for PC LFM
alpha_PC = (np.eye(n_) - beta_PC @ gamma_PC.T) @ m_x
# Step 7: Compute the covariance of residuals
s2_PC = sig @ e[:, k_:n_] * lambda2[k_:n_] @ e[:, k_:n_].T @ sig.T
return alpha_PC, beta_PC, gamma_PC, s2_PC
def fit_lfm_roblasso(x,
z,
p=None,
nu=1e9,
lambda_beta=0.,
lambda_phi=0.,
tol=1e-3,
fit_intercept=True,
maxiter=500,
print_iter=False,
rescale=False):
"""For details, see here.
Parameters
----------
x : array, shape(t_, n_)
z : array, shape(t_, k_)
p : array, optional, shape(t_)
nu : scalar, optional
lambda_beta : scalar, optional
lambda_phil : scalar, optional
tol : float, optional
fit_intercept: bool, optional
maxiter : scalar, optional
print_iter : bool, optional
rescale : bool, optional
Returns
-------
alpha_RMLFP : array, shape(n_,)
beta_RMLFP : array, shape(n_,k_)
sig2_RMLFP : array, shape(n_,n_)
"""
if len(x.shape) == 1:
x = x.reshape(-1, 1)
if len(z.shape) == 1:
z = z.reshape(-1, 1)
t_, n_ = x.shape
if p is None:
p = np.ones(t_) / t_
# rescale the variables
if rescale is True:
_, sigma2_x = meancov_sp(x, p)
sigma_x = np.sqrt(np.diag(sigma2_x))
x = x / sigma_x
_, sigma2_z = meancov_sp(z, p)
sigma_z = np.sqrt(np.diag(sigma2_z))
z = z / sigma_z
# Step 0: Set initial values using method of moments
alpha, beta, sigma2, u = fit_lfm_lasso(x,
z,
p,
lambda_beta,
fit_intercept=fit_intercept)
mu_u = np.zeros(n_)
for i in range(maxiter):
# Step 1: Update the weights
if nu >= 1e3 and np.linalg.det(sigma2) < 1e-13:
w = np.ones(t_)
else:
w = (nu + n_) / (nu + mahalanobis_dist(u, mu_u, sigma2)**2)
q = w * p
q = q / np.sum(q)
# Step 2: Update location and dispersion parameters
alpha_old, beta_old = alpha, beta
alpha, beta, sigma2, u = fit_lfm_lasso(x,
z,
q,
lambda_beta,
fit_intercept=fit_intercept)
sigma2, _ = graphical_lasso((w @ p) * sigma2, lambda_phi)
# Step 3: Check convergence
errors = [
np.linalg.norm(alpha - alpha_old, ord=np.inf) /
max(np.linalg.norm(alpha_old, ord=np.inf), 1e-20),
np.linalg.norm(beta - beta_old, ord=np.inf) /
max(np.linalg.norm(beta_old, ord=np.inf), 1e-20)
]
# print the loglikelihood and the error
if print_iter is True:
print('Iter: %i; Loglikelihood: %.5f; Errors: %.5f' %
(i, p @ mvt_logpdf(u, mu_u, sigma2, nu) -
lambda_beta * np.linalg.norm(beta, ord=1), max(errors)))
if max(errors) <= tol:
break
if rescale is True:
alpha = alpha * sigma_x
beta = ((beta / sigma_z).T * sigma_x).T
sigma2 = (sigma2.T * sigma_x).T * sigma_x
return alpha, beta, sigma2
def fit_lfm_mlfp(x,
z,
p=None,
nu=4,
tol=1e-3,
fit_intercept=True,
maxiter=500,
print_iter=False,
rescale=False):
"""For details, see here.
Parameters
----------
x : array, shape (t_, n_) if n_>1 or (t_, ) for n_=1
z : array, shape (t_, k_) if k_>1 or (t_, ) for k_=1
p : array, optional, shape (t_,)
nu : scalar, optional
tol : float, optional
fit_intercept: bool, optional
maxiter : scalar, optional
print_iter : bool, optional
rescale : bool, optional
Returns
-------
alpha : array, shape (n_,)
beta : array, shape (n_, k_) if k_>1 or (n_, ) for k_=1
sigma2 : array, shape (n_, n_)
u : shape (t_, n_) if n_>1 or (t_, ) for n_=1
"""
if np.ndim(x) < 2:
x = x.reshape(-1, 1).copy()
t_, n_ = x.shape
if np.ndim(z) < 2:
z = z.reshape(-1, 1).copy()
t_, n_ = x.shape
k_ = z.shape[1]
if p is None:
p = np.ones(t_) / t_
# rescale the variables
if rescale is True:
_, sigma2_x = meancov_sp(x, p)
sigma_x = np.sqrt(np.diag(sigma2_x))
x = x.copy() / sigma_x
_, sigma2_z = meancov_sp(z, p)
sigma_z = np.sqrt(np.diag(sigma2_z))
z = z.copy() / sigma_z
# Step 0: Set initial values using method of moments
alpha, beta, sigma2, u = fit_lfm_ols(x, z, p, fit_intercept=fit_intercept)
alpha, beta, sigma2, u = \
alpha.reshape((n_, 1)), beta.reshape((n_, k_)), \
sigma2.reshape((n_, n_)), u.reshape((t_, n_))
if nu > 2.:
# if nu <=2, then the covariance is not defined
sigma2 = (nu - 2.) / nu * sigma2
mu_u = np.zeros(n_)
for i in range(maxiter):
# Step 1: Update the weights and historical flexible probabilities
if nu >= 1e3 and np.linalg.det(sigma2) < 1e-13:
w = np.ones(t_)
else:
w = (nu + n_) / (nu + mahalanobis_dist(u, mu_u, sigma2)**2)
q = w * p
q = q / np.sum(q)
# Step 2: Update shift parameters, factor loadings and covariance
alpha_old, beta_old, sigma2_old = alpha, beta, sigma2
alpha, beta, sigma2, u = fit_lfm_ols(x,
z,
q,
fit_intercept=fit_intercept)
alpha, beta, sigma2, u = \
alpha.reshape((n_, 1)), beta.reshape((n_, k_)), \
sigma2.reshape((n_, n_)), u.reshape((t_, n_))
sigma2 = (w @ p) * sigma2
# Step 3: Check convergence
beta_tilde_old = np.column_stack((alpha_old, beta_old))
beta_tilde = np.column_stack((alpha, beta))
errors = [
np.linalg.norm(beta_tilde - beta_tilde_old, ord=np.inf) /
np.linalg.norm(beta_tilde_old, ord=np.inf),
np.linalg.norm(sigma2 - sigma2_old, ord=np.inf) /
np.linalg.norm(sigma2_old, ord=np.inf)
]
# print the loglikelihood and the error
if print_iter is True:
print('Iter: %i; Loglikelihood: %.5f; Errors: %.3e' %
(i, p @ mvt_logpdf(u, mu_u, sigma2, nu), max(errors)))
if max(errors) < tol:
break
if rescale is True:
alpha = alpha * sigma_x
beta = ((beta / sigma_z).T * sigma_x).T
sigma2 = (sigma2.T * sigma_x).T * sigma_x
return np.squeeze(alpha), np.squeeze(beta), np.squeeze(sigma2), np.squeeze(
u)
def invariance_test_ellipsoid(epsi,
l_,
*,
conf_lev=0.95,
fit=0,
r=2,
title='Invariance test',
bl=None,
bu=None,
plot_test=True):
"""For details, see here.
Parameters
----------
epsi : array, shape(t_)
l_ : scalar
conf_lev : scalar, optional
fit : scalar, optional
r : scalar, optional
title : string, optional
bl : scalar, optional
bu : scalar, optional
plot_test : boolean, optional
Returns
-------
rho : array, shape(l_)
conf_int : array, shape(2)
"""
if len(epsi.shape) == 2:
epsi = epsi.reshape(-1)
if bl is None:
bl = np.percentile(epsi, 0.25)
if bu is None:
bu = np.percentile(epsi, 99.75)
# Settings
np.seterr(invalid='ignore')
sns.set_style('white')
nb = int(np.round(10 *
np.log(epsi.shape))) # number of bins for histograms
# Step 1: compute the sample autocorrelations
rho = np.array([
st.pearsonr(epsi[:-k] - meancov_sp(epsi[:-k])[0],
epsi[k:] - meancov_sp(epsi[k:])[0])[0]
for k in range(1, l_ + 1)
])
# Step 2: compute confidence interval
alpha = 1 - conf_lev
z_alpha_half = st.norm.ppf(1 - alpha / 2) / np.sqrt(epsi.shape[0])
conf_int = np.array([-z_alpha_half, z_alpha_half])
# Step 3: plot the ellipse, if requested
if plot_test:
plt.style.use('arpm')
# Ellipsoid test: location-dispersion parameters
x = epsi[:-l_]
epsi = epsi[l_:]
z = np.concatenate((x.reshape((-1, 1)), epsi.reshape((-1, 1))), axis=1)
# Step 3: Compute the sample mean and sample covariance and generate figure
mu_hat, sigma2_hat = meancov_sp(z)
f = plt.figure()
f.set_size_inches(16, 9)
gs = plt.GridSpec(9, 16, hspace=1.2, wspace=1.2)
# max and min value of the first reference axis settings,
# for the scatter and histogram plots
# scatter plot (with ellipsoid)
xx = x.copy()
yy = epsi.copy()
xx[x < bl] = np.NaN
xx[x > bu] = np.NaN
yy[epsi < bl] = np.NaN
yy[epsi > bu] = np.NaN
ax_scatter = f.add_subplot(gs[1:6, 4:9])
ax_scatter.scatter(xx, yy, marker='.', s=10)
ax_scatter.ticklabel_format(axis='x', style='sci', scilimits=(-2, 2))
plt.xlabel('obs', fontsize=17)
plt.ylabel('lagged obs.', fontsize=17)
plot_ellipse(mu_hat,
sigma2_hat,
r=r,
plot_axes=False,
plot_tang_box=False,
color='orange',
line_width=2,
display_ellipse=True)
plt.suptitle(title, fontsize=20)
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
ax_scatter.set_xlim(np.array([bl, bu]))
ax_scatter.set_ylim(np.array([bl, bu]))
ax = f.add_subplot(gs[7:, 4:9])
# histogram plot of observations
xxx = x[~np.isnan(xx)]
px = np.ones(xxx.shape[0]) / xxx.shape[0]
nx, cx = histogram_sp(xxx, p=px, k_=nb)
hist_kws = {'weights': px.flatten(), 'edgecolor': 'k'}
fit_kws = {'color': 'orange', 'cut': 0}
if fit == 1: # normal
sns.distplot(xxx, hist_kws=hist_kws, kde=False, fit=st.norm, ax=ax)
plt.legend(['Normal fit', 'Marginal distr'], fontsize=14)
elif fit == 2 and sum(x < 0) == 0: # exponential
sns.distplot(xxx,
hist_kws=hist_kws,
fit_kws=fit_kws,
kde=False,
fit=st.expon,
ax=ax)
plt.legend(['Exponential fit', 'Marginal distr'], fontsize=14)
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
elif fit == 3 and sum(x < 0) == 0: # Poisson
ax.bar(cx,
nx,
cx[1] - cx[0],
facecolor=[0.8, 0.8, 0.8],
edgecolor='k')
k = np.arange(x.max() + 1)
mlest = x.mean()
plt.plot(k,
st.poisson.pmf(k, mlest),
'o',
linestyle='-',
lw=1,
markersize=3,
color='orange')
plt.legend(['Poisson fit', 'Marginal distr.'], loc=1, fontsize=14)
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
else:
ax.bar(cx,
nx,
cx[1] - cx[0],
facecolor=[0.8, 0.8, 0.8],
edgecolor='k')
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
ax.get_xaxis().set_visible(False)
ax.set_xlim(np.array([bl, bu]))
ax.spines['top'].set_visible(False)
ax.spines['right'].set_visible(False)
ax.invert_yaxis()
ax = f.add_subplot(gs[1:6, 0:3])
# histogram plot of lagged observations
yyy = epsi[~np.isnan(yy)]
py = np.ones(yyy.shape[0]) / yyy.shape[0]
hist_kws = {'weights': py.flatten(), 'edgecolor': 'k'}
fit_kws = {'color': 'orange', 'cut': 0}
ny, cy = histogram_sp(yyy, p=py, k_=nb)
if fit == 1:
sns.distplot(yyy,
hist_kws=hist_kws,
kde=False,
fit=st.norm,
vertical=True,
ax=ax)
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
elif fit == 2 and sum(epsi < 0) == 0:
sns.distplot(yyy,
hist_kws=hist_kws,
fit_kws=fit_kws,
kde=False,
fit=st.expon,
vertical=True,
ax=ax)
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
elif fit == 3 and sum(epsi < 0) == 0:
ax.barh(cy,
ny,
cy[1] - cy[0],
facecolor=[0.8, 0.8, 0.8],
edgecolor='k')
mlest = epsi.mean()
k = np.arange(epsi.max() + 1)
plt.plot(st.poisson.pmf(k, mlest),
k,
'o',
linestyle='-',
lw=1,
markersize=3,
color='orange')
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
else:
ax.barh(cy,
ny,
cy[1] - cy[0],
facecolor=[0.8, 0.8, 0.8],
edgecolor='k')
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
ax.get_yaxis().set_visible(False)
ax.spines['top'].set_visible(False)
ax.spines['right'].set_visible(False)
ax.set_ylim(np.array([bl, bu]))
ax.invert_xaxis()
# autocorrelation plot
ax = f.add_subplot(gs[1:6, 10:])
xx = np.arange(1, l_ + 1)
xxticks = xx
if len(xx) > 15:
xxticks = np.linspace(1, l_ + 1, 10, dtype=int)
plt.bar(xx, rho[:l_], 0.5, facecolor=[.8, .8, .8], edgecolor='k')
plt.bar(xx[l_ - 1],
rho[l_ - 1],
0.5,
facecolor='orange',
edgecolor='k') # highlighting the last bar
plt.plot([0, xx[-1] + 1], [conf_int[0], conf_int[0]], ':k')
plt.plot([0, xx[-1] + 1], [-conf_int[0], -conf_int[0]], ':k')
plt.xlabel('lag', fontsize=17)
plt.ylabel('Autocorrelation', fontsize=17)
plt.axis([0.5, l_ + 0.5, -1, 1])
plt.xticks(xxticks)
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
plt.grid(False)
ax.spines['top'].set_visible(False)
ax.spines['right'].set_visible(False)
return rho, conf_int
def fit_locdisp_mlfp(epsi,
*,
p=None,
nu=1000,
threshold=1e-3,
maxiter=1000,
print_iter=False):
"""For details, see here.
Parameters
----------
epsi : array, shape (t_, i_)
p : array, shape (t_,), optional
nu: float, optional
threshold : float, optional
maxiter : int, optional
print_iter : bool
Returns
-------
mu : array, shape (i_,)
sigma2 : array, shape (i_, i_)
"""
if len(epsi.shape) == 1:
epsi = epsi.reshape(-1, 1)
t_, i_ = epsi.shape
if p is None:
p = np.ones(t_) / t_
# Step 0: Set initial values using method of moments
mu, sigma2 = meancov_sp(epsi, p)
if nu > 2.:
# if nu <=2, then the covariance is not defined
sigma2 = (nu - 2.) / nu * sigma2
for i in range(maxiter):
# Step 1: Update the weights
if nu >= 1e3 and np.linalg.det(sigma2) < 1e-13:
w = np.ones(t_)
else:
w = (nu + i_) / (nu + mahalanobis_dist(epsi, mu, sigma2)**2)
q = w * p
# Step 2: Update location and dispersion parameters
mu_old, sigma2_old = mu, sigma2
mu, sigma2 = meancov_sp(epsi, q)
mu = mu / np.sum(q)
# Step 3: Check convergence
err = max(
np.linalg.norm(mu - mu_old, ord=np.inf) /
np.linalg.norm(mu_old, ord=np.inf),
np.linalg.norm(sigma2 - sigma2_old, ord=np.inf) /
np.linalg.norm(sigma2_old, ord=np.inf))
if print_iter is True:
print('Iter: %i; Loglikelihood: %.5f; Error: %.5f' %
(i, p @ mvt_logpdf(epsi, mu, sigma2, nu), err))
if err <= threshold:
break
return np.squeeze(mu), np.squeeze(sigma2)
def fit_garch_fp(dx,
p=None,
sigma2_0=None,
param0=None,
g=0.95,
rescale=False):
"""For details, see here.
Parameters
----------
dx : array, shape(t_,)
p : array, optional, shape(t_)
sigma2_0 : scalar, optional
param0 : list or array, optional, shape(4,)
g : scalar, optional
rescale : bool, optional
Returns
-------
param : list
sigma2 : array, shape(t_,)
epsi : array, shape(t_,)
"""
t_ = dx.shape[0]
# flexible probabilities
if p is None:
p = np.ones(t_) / t_
# sample mean and variance
if (param0 is None) or (rescale is True):
m, s2 = meancov_sp(dx, p)
if param0 is None:
param0 = [0.01, g - 0.01, s2 * (1. - g), m] # initial parameters
# Step 0: Set default standard variance if not provided
if sigma2_0 is None:
p_tau = exp_decay_fp(t_, t_ / 3, 0)
_, sigma2_0 = meancov_sp(dx, p_tau)
# Step 1: standardize data if requested
if rescale is True:
dx = (dx - m) / np.sqrt(s2)
param0[2] = param0[2] / s2
param0[3] = param0[3] - m
sigma2_0 = sigma2_0 / s2
# Step 2: Compute negative log-likelihood of GARCH
def theta(param):
a, b, c, mu = param
sigma2 = sigma2_0
theta = 0.0
for t in range(t_):
# if statement added because of overflow when sigma2 is too low
if np.abs(sigma2) > 1e-128:
theta = theta + (
(dx[t] - mu)**2 / sigma2 + np.log(sigma2)) * p[t]
sigma2 = c + a * (dx[t] - mu)**2 + b * sigma2
return theta
# Step 3: Minimize the negative log-likelihood
# parameter boundaries
bnds = ((1e-20, 1.), (1e-20, 1.), (1e-20, None), (None, None))
# stationary constraints
cons = {'type': 'ineq', 'fun': lambda param: g - param[0] - param[1]}
a_hat, b_hat, c_hat, mu_hat = \
minimize(theta, param0, bounds=bnds, constraints=cons)['x']
# Step 4: Compute realized variance and invariants
sigma2_hat = np.full(t_, sigma2_0)
for t in range(t_ - 1):
sigma2_hat[t + 1] = c_hat + a_hat * (dx[t] - mu_hat) ** 2 + \
b_hat * sigma2_hat[t]
epsi = (dx - mu_hat) / np.sqrt(sigma2_hat)
# Step 5: revert standardization at Step 1, if requested
if rescale is True:
c_hat = c_hat * s2
mu_hat = mu_hat * np.sqrt(s2) + m
sigma2_hat = sigma2_hat * s2
return np.array([a_hat, b_hat, c_hat,
mu_hat]), sigma2_hat, np.squeeze(epsi)
def cointegration_fp(x, p=None, *, b_threshold=0.99):
"""For details, see here.
Parameters
----------
x : array, shape(t_, d_)
p : array, shape(t_, d_)
b_threshold : scalar
Returns
-------
c_hat : array, shape(d_, l_)
beta_hat : array, shape(l_, )
"""
t_ = x.shape[0]
if len(x.shape) == 1:
x = x.reshape((t_, 1))
d_ = 1
else:
d_ = x.shape[1]
if p is None:
p = np.ones(t_) / t_
if p is None:
p = np.ones(t_)/t_
# Step 1: estimate HFP covariance matrix
_, sigma2_hat = meancov_sp(x, p)
# Step 2: find eigenvectors
e_hat, _ = pca_cov(sigma2_hat)
# Step 3: detect cointegration vectors
c_hat = []
b_hat = []
p = p[:-1]
for d in np.arange(0, d_):
# Step 4: Define series
y_t = e_hat[:, d] @ x.T
# Step 5: fit AR(1)
yt = y_t[1:].reshape((-1, 1))
ytm1 = y_t[:-1].reshape((-1, 1))
_, b, _, _ = fit_lfm_mlfp(yt, ytm1, p / np.sum(p))
if np.ndim(b) < 2:
b = np.array(b).reshape(-1, 1)
# Step 6: check stationarity
if abs(b[0, 0]) <= b_threshold:
c_hat.append(list(e_hat[:, d]))
b_hat.append(b[0, 0])
# Output
c_hat = np.array(c_hat).T
b_hat = np.array(b_hat)
# Step 7: Sort according to the AR(1) parameters beta_hat
c_hat = c_hat[:, np.argsort(b_hat)]
b_hat = np.sort(b_hat)
return c_hat, b_hat
def fit_lfm_ols(x_t, z_t, p_t=None, fit_intercept=True):
"""For details, see here.
Parameters
----------
x_t : array, shape (t_, n_) if n_>1 or (t_, ) for n_=1
z_t : array, shape (t_, k_) if k_>1 or (t_, ) for k_=1
p_t : array, optional, shape (t_,)
fit_intercept : bool
Returns
-------
alpha_hat_olsfp : array, shape (n_,)
beta_hat_olsfp : array, shape (n_, k_) if k_>1 or (n_, ) for k_=1
s2_u_hat_olsfp : array, shape (n_, n_)
u_t : array, shape (t_, n_) if n_>1 or (t_, ) for n_=1
"""
t_ = x_t.shape[0]
if len(z_t.shape) < 2:
z_t = z_t.reshape((t_, 1)).copy()
k_ = 1
else:
k_ = z_t.shape[1]
if len(x_t.shape) < 2:
x_t = x_t.reshape((t_, 1)).copy()
n_ = 1
else:
n_ = x_t.shape[1]
if p_t is None:
p_t = np.ones(t_) / t_
# Step 1: Compute HFP mean and covariance of (X,Z)'
if fit_intercept is True:
m_xz_hat_hfp, s2_xz_hat_hfp = meancov_sp(np.c_[x_t, z_t], p_t)
else:
m_xz_hat_hfp = np.zeros(n_ + k_)
s2_xz_hat_hfp = p_t * np.c_[x_t, z_t].T @ np.c_[x_t, z_t]
# Step 2: Compute the OLSFP estimates
s2_z_hat_hfp = s2_xz_hat_hfp[n_:, n_:]
s_x_z_hat_hfp = s2_xz_hat_hfp[:n_, n_:]
m_xz_hat_hfp = m_xz_hat_hfp.reshape(-1)
m_z_hat_hfp = m_xz_hat_hfp[n_:].reshape(-1, 1)
m_x_hat_hfp = m_xz_hat_hfp[:n_].reshape(-1, 1)
beta_hat_olsfp = s_x_z_hat_hfp @ np.linalg.inv(s2_z_hat_hfp)
alpha_hat_olsfp = m_x_hat_hfp - beta_hat_olsfp @ m_z_hat_hfp
# Step 3: Compute residuals and OLSFP estimate of covariance of U
u_t = (x_t.T - alpha_hat_olsfp - beta_hat_olsfp @ z_t.T).T
_, s2_u_hat_olsfp = meancov_sp(u_t, p_t)
return alpha_hat_olsfp[:, 0], np.squeeze(beta_hat_olsfp),\
np.squeeze(s2_u_hat_olsfp), np.squeeze(u_t)
def fit_dcc_t(dx, p=None, *, rho2=None, param0=None, g=0.99):
"""For details, see here.
Parameters
----------
dx : array, shape(t_, i_)
p : array, optional, shape(t_)
rho2 : array, shape(i_, i_)
param0 : list or array, shape(2,)
g : scalar, optional
Returns
-------
params : list, shape(3,)
r2_t : array, shape(t_, i_, i_)
epsi : array, shape(t_, i_)
q2_t_ : array, shape(i_, i_)
"""
# Step 0: Setup default values
t_, i_ = dx.shape
# flexible probabilities
if p is None:
p = np.ones(t_) / t_
# target correlation
if rho2 is None:
_, rho2 = meancov_sp(dx, p)
rho2, _ = cov_2_corr(rho2)
# initial parameters
if param0 is None:
param0 = [0.01, g - 0.01] # initial parameters
# Step 1: Compute negative log-likelihood of GARCH
def llh(params):
a, b = params
mu = np.zeros(i_)
q2_t = rho2.copy()
r2_t, _ = cov_2_corr(q2_t)
llh = 0.0
for t in range(t_):
llh = llh - p[t] * multivariate_normal.logpdf(dx[t, :], mu, r2_t)
q2_t = rho2 * (1 - a - b) + \
a * np.outer(dx[t, :], dx[t, :]) + b * q2_t
r2_t, _ = cov_2_corr(q2_t)
return llh
# Step 2: Minimize the negative log-likelihood
# parameter boundaries
bnds = ((1e-20, 1.), (1e-20, 1.))
# stationary constraints
cons = {'type': 'ineq', 'fun': lambda param: g - param[0] - param[1]}
a, b = minimize(llh, param0, bounds=bnds, constraints=cons)['x']
# Step 3: Compute realized correlations and residuals
q2_t = rho2.copy()
r2_t = np.zeros((t_, i_, i_))
r2_t[0, :, :], _ = cov_2_corr(q2_t)
for t in range(t_ - 1):
q2_t = rho2 * (1 - a - b) + \
a * np.outer(dx[t, :], dx[t, :]) + b * q2_t
r2_t[t + 1, :, :], _ = cov_2_corr(q2_t)
l_t = np.linalg.cholesky(r2_t)
epsi = np.linalg.solve(l_t, dx)
return [1. - a - b, a, b], r2_t, epsi, q2_t
