def measure_likelihood(self, y: np.array, u: np.array) -> float:
# Time before update. (mus indexed at +1 for initial mu)
tm1 = self.mus.shape[Axis.time] - 2
(A, B, C, D, Q, R) = self.parameters(tm1)
(mu, V) = self.state(tm1)
(y_pred, mu_pred) = self.predict(A, B, C, D, mu, u)
V_pred = self.predict_covariance(A, V, Q)
# Update observation
self.observe(tm1 + 1, y, u)
# Update State
return KalmanFilter.update(self,
tm1 + 1,
mu_pred,
V_pred,
y,
y_pred,
compute_likelihood=True)
[0]])
init_V = np.eye(A.shape[0], A.shape[1])
return (A, B, C, D, Q, R), init_mu, init_V
# Split the data into first half and second half
(r, c, ts) = obs.shape
training_start = 0
test_start = 252
n_data_points = 252
test_end = test_start + n_data_points
training_ys = obs[:, :, training_start:test_start]
training_us = conds[:, :, training_start:test_start]
test_ys = obs[:, :, test_start:test_end]
test_us = conds[:, :, test_start:test_end]
(kf, ll_hists) = KalmanFilter.fit(training_ys, training_us, gen_params(1, 4), iters=4)
# Show Kalman Filter hist
# for l in ll_hists:
# print(l, sum(l))
# Get test values
kf.ys = obs[:, :, :test_end]
kf.us = conds[:, :, :test_end]
(y_pred, ll, _) = kf.filter()
y_pred = y_pred[:, :, test_start:test_end]
(x, y, pred_ts) = y_pred.shape
pred = pd.DataFrame(y_pred.reshape((x, pred_ts)).T, columns=['Predicted Prices'])
D = np.array([[0, 0]])
Q = np.array([[0.1, 0],
[0, 0.1]])
R = np.array([[0.1]])
params = (A, B, C, D, Q, R)
# Initial Values
init_mu = np.array([[1],
[2]])
init_V = np.array([[INIT_VAR, 0],[0, INIT_VAR]])
init_state = (init_mu, init_V)
# Data Generator
space_gen = KalmanSpaceGenerator(params, init_mu, init_V, 100)
y_0tT, u_0tT = space_gen.generate_data()
# Kalman Filter
kf = KalmanFilter(params, init_mu, init_V)
# Assign data
(ys, us) = line_gen(100)
kf.ys = y_0tT
kf.us = u_0tT
y_pred, ll, s = kf.smooth_filter()
l1, = plt.plot(y_0tT.flatten(), label='Kalman Space', color='blue')
l2, = plt.plot(y_pred.flatten(), label='Online Predictions', color='green')
l3, = plt.plot(y_0tT.flatten() - y_pred.flatten(), label='Error', color='red')
plt.legend(handles=[l1, l2])
plt.show()
return (A, B, C, D, Q, R), init_mu, init_V
# Split the data into first half and second half
(r, c, ts) = obs.shape
training_start = 0
test_start = 252
n_data_points = 252
test_end = test_start + n_data_points
training_ys = obs[:, :, training_start:test_start]
training_us = conds[:, :, training_start:test_start]
test_ys = obs[:, :, test_start:test_end]
test_us = conds[:, :, test_start:test_end]
(kf, ll_hists) = KalmanFilter.fit(training_ys,
training_us,
gen_params(1, 4),
iters=4)
# Show Kalman Filter hist
# for l in ll_hists:
# print(l, sum(l))
# Get test values
kf.ys = obs[:, :, :test_end]
kf.us = conds[:, :, :test_end]
(y_pred, ll, _) = kf.filter()
y_pred = y_pred[:, :, test_start:test_end]
(x, y, pred_ts) = y_pred.shape
pred = pd.DataFrame(y_pred.reshape((x, pred_ts)).T,
B = np.array([[0, 0], [0, 0]]) C = np.array([[1, 0]]) D = np.array([[0, 0]]) Q = np.array([[0.1, 0], [0, 0.1]]) R = np.array([[0.1]]) params = (A, B, C, D, Q, R) # Initial Values init_mu = np.array([[1], [2]]) init_V = np.array([[INIT_VAR, 0], [0, INIT_VAR]]) init_state = (init_mu, init_V) # Data Generator space_gen = KalmanSpaceGenerator(params, init_mu, init_V, 100) y_0tT, u_0tT = space_gen.generate_data() # Kalman Filter kf = KalmanFilter(params, init_mu, init_V) # Assign data (ys, us) = line_gen(100) kf.ys = y_0tT kf.us = u_0tT y_pred, ll, s = kf.smooth_filter() l1, = plt.plot(y_0tT.flatten(), label='Kalman Space', color='blue') l2, = plt.plot(y_pred.flatten(), label='Online Predictions', color='green') l3, = plt.plot(y_0tT.flatten() - y_pred.flatten(), label='Error', color='red') plt.legend(handles=[l1, l2]) plt.show()
def observe(self, t, y_t, u_t):
KalmanFilter.observe(self, t, y_t, u_t)
def project(self, n):
return KalmanFilter.project(self, n)
def __init__(self, init_params, init_mu, init_V):
Particle.__init__(self)
KalmanFilter.__init__(self, init_params, init_mu, init_V)
[0, 0, 0, 0, 0], [0, 0, 0, 0, 0]])
C = np.array([[1, 0, 0, 0, 0], [0, 1, 0, 0, 0]])
D = np.array([[0, 0, 0, 0, 0], [0, 0, 0, 0, 0]])
Q = np.array([[0.8817, 0, 0, 0, 0], [0, 17.81, 0, 0, 0], [0, 0, 1, 0, 0],
[0, 0, 0, 0, 0], [0, 0, 0, 0, 1]])
R = np.array([[1, 0], [0, 1]])
init_mu = np.array([[0], [5], [0], [0], [0]])
init_V = np.array(
[[100000000000, 100000000000, 100000000000, 100000000000, 100000000000],
[100000000000, 100000000000, 100000000000, 100000000000, 100000000000],
[100000000000, 100000000000, 100000000000, 100000000000, 100000000000],
[100000000000, 100000000000, 100000000000, 100000000000, 100000000000],
[100000000000, 100000000000, 100000000000, 100000000000, 100000000000]])
kf = KalmanFilter((A, B, C, D, Q, R), init_mu, init_V)
kf.ys = obs
kf.us = conds
(y_pred, ll, _) = kf.smooth_filter()
(x, y, t) = y_pred.shape
pred = pd.DataFrame(y_pred.reshape((x, t)).T,
columns=['Predicted Liquidity', 'Predicted Prices'])
# Projected prices
# (projection, V_projections) = kf.project(30)
# plt.plot(projection[1, :, :].flatten(), c='r')
# plt.show()
# Ordinary Least Squares Calculation
ols = sum((rmd.data.values[:, 2] - pred.values[:, 1]) *
[0, 0, 0, 0, 1]])
R = np.array([[1, 0],
[0, 1]])
init_mu = np.array([[0],
[5],
[0],
[0],
[0]])
init_V = np.array([[100000000000, 100000000000, 100000000000, 100000000000, 100000000000],
[100000000000, 100000000000, 100000000000, 100000000000, 100000000000],
[100000000000, 100000000000, 100000000000, 100000000000, 100000000000],
[100000000000, 100000000000, 100000000000, 100000000000, 100000000000],
[100000000000, 100000000000, 100000000000, 100000000000, 100000000000]])
kf = KalmanFilter((A, B, C, D, Q, R), init_mu, init_V)
kf.ys = obs
kf.us = conds
(y_pred, ll, _) = kf.smooth_filter()
(x, y, t) = y_pred.shape
pred = pd.DataFrame(y_pred.reshape((x, t)).T, columns=['Predicted Liquidity', 'Predicted Prices'])
# Projected prices
# (projection, V_projections) = kf.project(30)
# plt.plot(projection[1, :, :].flatten(), c='r')
# plt.show()
# Ordinary Least Squares Calculation
ols = sum((rmd.data.values[:, 2] - pred.values[:, 1]) * (rmd.data.values[:, 2] - pred.values[:, 1])) / 2
print(ols)
# Check if we have already trained a model for this.
# TODO Move to a caching module, to hide the cache and restore kalman filters.
# TODO Simplify Caching code
possible_files = cache_utils.get_matching_ids("kalman", stock, training_start_dates[period],
test_start_dates[period], True, cache_dir=CACHE_DIR)
identified_cached_kalman = [kalman_identified(f, data_manager.features(), True, cache_dir=CACHE_DIR)
for f in possible_files]
if any(identified_cached_kalman):
# Debug purposes
if len(identified_cached_kalman) > 1:
raise Exception("There is a problem with the cache!")
# Get the identified:
kalman_files = [possible_files[i] for i, is_match in enumerate(identified_cached_kalman) if is_match]
cached_kalman = kalman_files.pop(0)
kf = KalmanFilter.restore(cached_kalman, cache_dir=CACHE_DIR)
print("RESTORED {cached}.".format(cached=cached_kalman))
else:
# Train a new kalman filter and cache it
kf = None
for n_its in range(8, 0, -1):
run_title = "{stock} from period {start} - {finish} with {n_its} EM iterations"\
.format(stock=stock, start=utils.pretty_date_str(training_start_dates[period]),
finish=utils.pretty_date_str(test_start_dates[period]), n_its=n_its)
kf = kalman_gym.select_best_model(training_ys, training_us, iters=n_its, min_fit=n_obs - 10,
n_models=25, run_title=run_title)
if kf is not None:
break
if kf is None: