def func(self, _params: np.ndarray):
# split params
if self.use_mu:
mu = _params[-1]
else:
mu = self.mu_arr
const, alpha, beta, _ = np.split(_params, self.split_index)
alpha = ilogit(alpha)
beta = ilogit(beta)
const = np.exp(const)
square_arr = (self.arr - mu)**2
arch_x_arr = stack_delay_arr(square_arr, self.alpha_num)
if self.beta_num > 0: # estimate latent_arr
self.latent_arr[self.beta_num:] = alpha.dot(arch_x_arr) + const
self.latent_arr[:self.beta_num] = square_arr.mean()
garch_recursion(self.latent_arr, beta)
else:
self.latent_arr = alpha.dot(arch_x_arr) + const
loss = -logpdf(self.y_arr, mu, self.latent_arr[self.beta_num:]).mean()
return loss
def func(self, _params: np.ndarray):
# split params
if self.use_mu:
mu = _params[-1]
else:
mu = self.mu_arr
const, phi, theta, alpha, beta, _ = np.split(_params, self.split_index)
const = np.exp(const)
alpha = ilogit(alpha)
beta = ilogit(beta)
# arma part
if self.phi_num > 0:
mu = phi.dot(self.x_arr) + mu
if self.theta_num > 0:
self.latent_arma_arr[:self.theta_num] = 0.0
self.latent_arma_arr[self.theta_num:] = self.y_arr - mu
arma_recursion(self.latent_arma_arr, theta)
else:
self.latent_arma_arr = self.y_arr - mu
# garch part
new_info_arr = self.latent_arma_arr[self.theta_num:]
square_arr = new_info_arr**2
arch_x_arr = stack_delay_arr(square_arr, self.alpha_num)
if self.beta_num > 0: # estimate latent_arr
self.latent_garch_arr[self.
beta_num:] = alpha.dot(arch_x_arr) + const
self.latent_garch_arr[:self.beta_num] = square_arr.mean()
garch_recursion(self.latent_garch_arr, beta)
else:
self.latent_garch_arr = alpha.dot(arch_x_arr) + const
loss = -logpdf(new_info_arr[self.alpha_num:], 0,
self.latent_garch_arr[self.beta_num:]).mean()
return loss
def optimize(
self,
_init_params: np.ndarray,
_max_iter: int = 200,
_disp: bool = False,
):
cons = {
'type':
'eq',
'fun':
lambda x: ilogit(x[self.split_index[2]:self.split_index[4]]).sum()
- 1.0,
}
ret = minimize(self.func,
_init_params,
constraints=cons,
method='SLSQP',
options={
'maxiter': _max_iter,
'disp': _disp
})
return ret.x
def optimize(
self,
_init_params: np.ndarray,
_max_iter: int = 50,
_disp: bool = False,
):
cons = {
'type':
'eq',
'fun':
lambda x: ilogit(x[1:self.alpha_num + self.beta_num + 1]).sum() -
1.0,
}
ret = minimize(self.func,
_init_params,
constraints=cons,
method='SLSQP',
options={
'maxiter': _max_iter,
'disp': _disp
})
return ret.x
def getBetas(self) -> np.ndarray:
return ilogit(self.logit_beta_arr)
def getAlphas(self) -> np.ndarray:
return ilogit(self.logit_alpha_arr)