def backward_sample(self, steps=1):
"""
Generate state sequence using distributions:
.. math:: p(\theta_{t} | \theta_{t + k} D_t)
"""
from statlib.distributions import rmvnorm
if steps != 1:
raise Exception('only do one step backward sampling for now...')
T = self.nobs
# Backward sample
mu_draws = np.zeros((T + 1, self.ndim))
m = self.mu_mode
C = self.mu_scale
a = self.mu_forc_mode
R = self.mu_forc_scale
mu_draws[T] = rmvnorm(m[T], C[T])
# initial values for smoothed dist'n
for t in xrange(T-1, -1, -1):
# B_{t} = C_t G_t+1' R_t+1^-1
B = chain_dot(C[t], self.G.T, LA.inv(R[t+1]))
# smoothed mean
ht = m[t] + np.dot(B, mu_draws[t+1] - a[t+1])
Ht = C[t] - chain_dot(B, R[t+1], B.T)
mu_draws[t] = rmvnorm(ht, np.atleast_2d(Ht))
return mu_draws.squeeze()
def backward_sample(self, steps=1):
"""
Generate state sequence using distributions:
.. math:: p(\theta_{t} | \theta_{t + k} D_t)
"""
from statlib.distributions import rmvnorm
if steps != 1:
raise Exception('only do one step backward sampling for now...')
T = self.nobs
# Backward sample
mu_draws = np.zeros((T + 1, self.ndim))
m = self.mu_mode
C = self.mu_scale
a = self.mu_forc_mode
R = self.mu_forc_scale
mu_draws[T] = rmvnorm(m[T], C[T])
# initial values for smoothed dist'n
for t in xrange(T - 1, -1, -1):
# B_{t} = C_t G_t+1' R_t+1^-1
B = chain_dot(C[t], self.G.T, LA.inv(R[t + 1]))
# smoothed mean
ht = m[t] + np.dot(B, mu_draws[t + 1] - a[t + 1])
Ht = C[t] - chain_dot(B, R[t + 1], B.T)
mu_draws[t] = rmvnorm(ht, np.atleast_2d(Ht))
return mu_draws.squeeze()
def _update_mu(v, w, phi, lam):
# FFBS
# allocate result arrays
mode = np.zeros((T + 1, p))
a = np.zeros((T + 1, p))
C = np.zeros((T + 1, p))
R = np.zeros((T + 1, p))
# simple priors...
mode[0] = 0
C[0] = np.eye(p)
# Forward filter
Ft = m_([[1]])
for i, obs in enumerate(y):
t = i + 1
at = phi * mode[t - 1] if t > 1 else mode[0]
Rt = phi**2 * C[t - 1] + w if t > 1 else C[0]
Vt = lam[t - 1] * v
Qt = chain_dot(Ft.T, Rt, Ft) + Vt
At = np.dot(Rt, Ft) / Qt
# forecast theta as time t
ft = np.dot(Ft.T, at)
err = obs - ft
# update mean parameters
mode[t] = at + np.dot(At, err)
C[t] = Rt - np.dot(At, np.dot(Qt, At.T))
a[t] = at
R[t] = Rt
# Backward sample
mu = np.zeros((T + 1, p))
# initial values for smoothed dist'n
fR = C[-1]
fm = mode[-1]
for t in xrange(T + 1):
if t < T:
# B_{t} = C_t G_t+1' R_t+1^-1
B = np.dot(C[t] * phi, la.inv(np.atleast_2d(R[t + 1])))
# smoothed mean
fm = mode[t] + np.dot(B, mode[t + 1] - a[t + 1])
fR = C[t] + chain_dot(B, C[t + 1] - R[t + 1], B.T)
mu[t] = dist.rmvnorm(fm, np.atleast_2d(fR))
return mu.squeeze()
def _update_mu(v, w, phi, lam):
# FFBS
# allocate result arrays
mode = np.zeros((T + 1, p))
a = np.zeros((T + 1, p))
C = np.zeros((T + 1, p))
R = np.zeros((T + 1, p))
# simple priors...
mode[0] = 0
C[0] = np.eye(p)
# Forward filter
Ft = m_([[1]])
for i, obs in enumerate(y):
t = i + 1
at = phi * mode[t - 1] if t > 1 else mode[0]
Rt = phi ** 2 * C[t - 1] + w if t > 1 else C[0]
Vt = lam[t - 1] * v
Qt = chain_dot(Ft.T, Rt, Ft) + Vt
At = np.dot(Rt, Ft) / Qt
# forecast theta as time t
ft = np.dot(Ft.T, at)
err = obs - ft
# update mean parameters
mode[t] = at + np.dot(At, err)
C[t] = Rt - np.dot(At, np.dot(Qt, At.T))
a[t] = at
R[t] = Rt
# Backward sample
mu = np.zeros((T + 1, p))
# initial values for smoothed dist'n
fR = C[-1]
fm = mode[-1]
for t in xrange(T + 1):
if t < T:
# B_{t} = C_t G_t+1' R_t+1^-1
B = np.dot(C[t] * phi, la.inv(np.atleast_2d(R[t+1])))
# smoothed mean
fm = mode[t] + np.dot(B, mode[t+1] - a[t+1])
fR = C[t] + chain_dot(B, C[t+1] - R[t+1], B.T)
mu[t] = dist.rmvnorm(fm, np.atleast_2d(fR))
return mu.squeeze()
def backward_smooth(self):
"""
Compute posterior estimates of state vector given full data set,
i.e. p(\theta_t | D_T)
cf. W&H sections
4.7 / 4.8: regular smoothing recurrences (Theorem 4.4)
10.8.4 adjustments for variance discounting
Notes
-----
\theta_{t-k} | D_t ~ T_{n_t} [a_t(-k), (S_t / S_{t-k}) R_t(-k)]
Returns
-------
(filtered state mode,
filtered state cov,
filtered degrees of freedom)
"""
beta = self.var_discount
T = self.nobs
a = self.mu_forc_mode
R = self.mu_forc_scale
fdf = self.df.copy()
fS = self.var_est.copy()
fm = self.mu_mode.copy()
fC = self.mu_scale.copy()
C = fC[T]
for t in xrange(T - 1, -1, -1):
B = chain_dot(fC[t], self.G.T, LA.inv(R[t+1]))
# W&H p. 364
fdf[t] = (1 - beta) * fdf[t] + beta * fdf[t + 1]
fS[t] = 1 / ((1 - beta) / fS[t] + beta / fS[t+1])
# W&H p. 113
fm[t] = fm[t] + np.dot(B, fm[t+1] - a[t+1])
C = fC[t] + chain_dot(B, C - R[t+1], B.T)
fC[t] = C * fS[T] / fS[t]
return fm, fC, fdf
def backward_smooth(self):
"""
Compute posterior estimates of state vector given full data set,
i.e. p(\theta_t | D_T)
cf. W&H sections
4.7 / 4.8: regular smoothing recurrences (Theorem 4.4)
10.8.4 adjustments for variance discounting
Notes
-----
\theta_{t-k} | D_t ~ T_{n_t} [a_t(-k), (S_t / S_{t-k}) R_t(-k)]
Returns
-------
(filtered state mode,
filtered state cov,
filtered degrees of freedom)
"""
beta = self.var_discount
T = self.nobs
a = self.mu_forc_mode
R = self.mu_forc_scale
fdf = self.df.copy()
fS = self.var_est.copy()
fm = self.mu_mode.copy()
fC = self.mu_scale.copy()
C = fC[T]
for t in xrange(T - 1, -1, -1):
B = chain_dot(fC[t], self.G.T, LA.inv(R[t + 1]))
# W&H p. 364
fdf[t] = (1 - beta) * fdf[t] + beta * fdf[t + 1]
fS[t] = 1 / ((1 - beta) / fS[t] + beta / fS[t + 1])
# W&H p. 113
fm[t] = fm[t] + np.dot(B, fm[t + 1] - a[t + 1])
C = fC[t] + chain_dot(B, C - R[t + 1], B.T)
fC[t] = C * fS[T] / fS[t]
return fm, fC, fdf
def calc_update(jt, jtp):
model = self.models[jt]
G = model.G
prior_scale = self.mu_scale[t - 1, jtp]
if t > 1:
# only discount after first time step! hmm
Wt = model.get_Wt(prior_scale)
Rt = chain_dot(G, prior_scale, G.T) + Wt
else:
# use prior for t=1, because Mike does
Rt = prior_scale
return Rt
def calc_update(jt, jtp):
model = self.models[jt]
G = model.G
prior_scale = self.mu_scale[t - 1, jtp]
if t > 1:
# only discount after first time step! hmm
Wt = model.get_Wt(prior_scale)
Rt = chain_dot(G, prior_scale, G.T) + Wt
else:
# use prior for t=1, because Mike does
Rt = prior_scale
return Rt
def get_Wt(self, Cprior):
# Eq. 12.18
disc_var = np.diag(Cprior) * (1 / self.deltas - 1)
return chain_dot(self.G, np.diag(disc_var), self.G.T)
def calc_update(jt, jtp):
prior_obs_var = (self.var_est[t - 1, jtp] *
self.models[jt].obs_var_mult)
return chain_dot(Ft.T, Rt[jt, jtp], Ft) + prior_obs_var
def filter_python(Y, F, G, delta, beta, df0, v0, m0, C0):
"""
Univariate DLM update equations with unknown observation variance
delta : state discount
beta : variance discount
"""
# cdef:
# Py_ssize_t i, t, nobs, ndim
# ndarray[double_t, ndim=1] df, Q, S
# ndarray a, C, R, mode
# ndarray at, mt, Ft, At, Ct, Rt
# double_t obs, ft, e, nt, dt, St, Qt
nobs = len(Y)
ndim = len(G)
mode = nan_array(nobs + 1, ndim)
a = nan_array(nobs + 1, ndim)
C = nan_array(nobs + 1, ndim, ndim)
R = nan_array(nobs + 1, ndim, ndim)
S = nan_array(nobs + 1)
Q = nan_array(nobs)
df = nan_array(nobs + 1)
mode[0] = mt = m0
C[0] = Ct = C0
df[0] = nt = df0
S[0] = St = v0
dt = df0 * v0
# allocate result arrays
for i in range(nobs):
obs = Y[i]
# column vector, for W&H notational consistency
# Ft = F[i]
Ft = F[i:i+1].T
# advance index: y_1 through y_nobs, 0 is prior
t = i + 1
# derive innovation variance from discount factor
at = mt
Rt = Ct
if t > 1:
# only discount after first time step?
if G is not None:
at = np.dot(G, mt)
Rt = chain_dot(G, Ct, G.T) / delta
else:
Rt = Ct / delta
# Qt = chain_dot(Ft.T, Rt, Ft) + St
Qt = chain_dot(Ft.T, Rt, Ft) + St
At = np.dot(Rt, Ft) / Qt
# forecast theta as time t
ft = np.dot(Ft.T, at)
e = obs - ft
# update mean parameters
mode[t] = mt = at + np.dot(At, e)
dt = beta * dt + St * e * e / Qt
nt = beta * nt + 1
St = dt / nt
S[t] = St
Ct = (S[t] / S[t-1]) * (Rt - np.dot(At, At.T) * Qt)
Ct = (Ct + Ct.T) / 2 # symmetrize
df[t] = nt
Q[t-1] = Qt
C[t] = Ct
a[t] = at
R[t] = Rt
return mode, a, C, df, S, Q, R
def _mvfilter_python(Y, F, G, V, delta, beta, df0, v0, m0, C0):
"""
Matrix-variate DLM update equations
V : V_t sequence
"""
nobs, k = Y.shape
p = F.shape[1]
mode = nan_array(nobs + 1, p, k) # theta posterior mode
C = nan_array(nobs + 1, p, p) # theta posterior scale
a = nan_array(nobs + 1, p, k) # theta forecast mode
Q = nan_array(nobs) # variance multiplier term
D = nan_array(nobs + 1, k, k) # scale matrix
S = nan_array(nobs + 1, k, k) # covariance estimate D / n
df = nan_array(nobs + 1)
mode[0] = m0
C[0] = C0
df[0] = df0
S[0] = v0
Mt = m0
Ct = C0
n = df0
d = n + k - 1
Dt = D0
St = Dt / n
# allocate result arrays
for t in xrange(1, nobs + 1):
obs = Y[t - 1:t].T
Ft = F[t - 1:t].T
# derive innovation variance from discount factor
# only discount after first time step?
if G is not None:
at = np.dot(G, Mt)
Rt = chain_dot(G, Ct, G.T) / delta
else:
at = Mt
Rt = Ct / delta
et = obs - np.dot(at.T, Ft)
qt = chain_dot(Ft.T, Rt, Ft) + V
At = np.dot(Rt, Ft) / qt
# update mean parameters
n = beta * n
b = (n + k - 1) / d
n = n + 1
d = n + k - 1
Dt = b * Dt + np.dot(et, et.T) / qt
St = Dt / n
Mt = at + np.dot(At, et.T)
Ct = Rt - np.dot(At, At.T) * qt
C[t] = Ct; df[t] = n; S[t] = (St + St.T) / 2; mode[t] = Mt
D[t] = Dt; a[t] = at; Q[t-1] = qt
return mode, a, C, S, Q
def filter_python(Y, F, G, delta, beta, df0, v0, m0, C0):
"""
Univariate DLM update equations with unknown observation variance
delta : state discount
beta : variance discount
"""
# cdef:
# Py_ssize_t i, t, nobs, ndim
# ndarray[double_t, ndim=1] df, Q, S
# ndarray a, C, R, mode
# ndarray at, mt, Ft, At, Ct, Rt
# double_t obs, ft, e, nt, dt, St, Qt
nobs = len(Y)
ndim = len(G)
mode = nan_array(nobs + 1, ndim)
a = nan_array(nobs + 1, ndim)
C = nan_array(nobs + 1, ndim, ndim)
R = nan_array(nobs + 1, ndim, ndim)
S = nan_array(nobs + 1)
Q = nan_array(nobs)
df = nan_array(nobs + 1)
mode[0] = mt = m0
C[0] = Ct = C0
df[0] = nt = df0
S[0] = St = v0
dt = df0 * v0
# allocate result arrays
for i in range(nobs):
obs = Y[i]
# column vector, for W&H notational consistency
# Ft = F[i]
Ft = F[i:i + 1].T
# advance index: y_1 through y_nobs, 0 is prior
t = i + 1
# derive innovation variance from discount factor
at = mt
Rt = Ct
if t > 1:
# only discount after first time step?
if G is not None:
at = np.dot(G, mt)
Rt = chain_dot(G, Ct, G.T) / delta
else:
Rt = Ct / delta
# Qt = chain_dot(Ft.T, Rt, Ft) + St
Qt = chain_dot(Ft.T, Rt, Ft) + St
At = np.dot(Rt, Ft) / Qt
# forecast theta as time t
ft = np.dot(Ft.T, at)
e = obs - ft
# update mean parameters
mode[t] = mt = at + np.dot(At, e)
dt = beta * dt + St * e * e / Qt
nt = beta * nt + 1
St = dt / nt
S[t] = St
Ct = (S[t] / S[t - 1]) * (Rt - np.dot(At, At.T) * Qt)
Ct = (Ct + Ct.T) / 2 # symmetrize
df[t] = nt
Q[t - 1] = Qt
C[t] = Ct
a[t] = at
R[t] = Rt
return mode, a, C, df, S, Q, R
def _mvfilter_python(Y, F, G, V, delta, beta, df0, v0, m0, C0):
"""
Matrix-variate DLM update equations
V : V_t sequence
"""
nobs, k = Y.shape
p = F.shape[1]
mode = nan_array(nobs + 1, p, k) # theta posterior mode
C = nan_array(nobs + 1, p, p) # theta posterior scale
a = nan_array(nobs + 1, p, k) # theta forecast mode
Q = nan_array(nobs) # variance multiplier term
D = nan_array(nobs + 1, k, k) # scale matrix
S = nan_array(nobs + 1, k, k) # covariance estimate D / n
df = nan_array(nobs + 1)
mode[0] = m0
C[0] = C0
df[0] = df0
S[0] = v0
Mt = m0
Ct = C0
n = df0
d = n + k - 1
Dt = D0
St = Dt / n
# allocate result arrays
for t in xrange(1, nobs + 1):
obs = Y[t - 1:t].T
Ft = F[t - 1:t].T
# derive innovation variance from discount factor
# only discount after first time step?
if G is not None:
at = np.dot(G, Mt)
Rt = chain_dot(G, Ct, G.T) / delta
else:
at = Mt
Rt = Ct / delta
et = obs - np.dot(at.T, Ft)
qt = chain_dot(Ft.T, Rt, Ft) + V
At = np.dot(Rt, Ft) / qt
# update mean parameters
n = beta * n
b = (n + k - 1) / d
n = n + 1
d = n + k - 1
Dt = b * Dt + np.dot(et, et.T) / qt
St = Dt / n
Mt = at + np.dot(At, et.T)
Ct = Rt - np.dot(At, At.T) * qt
C[t] = Ct
df[t] = n
S[t] = (St + St.T) / 2
mode[t] = Mt
D[t] = Dt
a[t] = at
Q[t - 1] = qt
return mode, a, C, S, Q
def get_Wt(self, Cprior):
# Eq. 12.18
disc_var = np.diag(Cprior) * (1 / self.deltas - 1)
return chain_dot(self.G, np.diag(disc_var), self.G.T)
def calc_update(jt, jtp):
prior_obs_var = (self.var_est[t - 1, jtp] *
self.models[jt].obs_var_mult)
return chain_dot(Ft.T, Rt[jt, jtp], Ft) + prior_obs_var
