def expected_time_of_next_value(self, lagged_times):
"""
Rudimentary estimate of the next time a data point will arrive based on history
:returns float, float expected_at epoch time
dt typical time between data arrivals
"""
if len(lagged_times) > 5:
# Compute or assume time between data points
neg_dt = approx_mode(np.diff(lagged_times))
dt = -neg_dt if neg_dt is not None else None
if dt is None:
dt = self.DELAYS[-1]
if dt < 0:
pass
if dt < 10:
dt = 10
# Don't wait more than 6 hours
if dt > (60 * 60 * 6):
dt = 60 * 60 * 6
# Compute expected time of next data point, but if data is missing move it forward
# num_intervals>1 is used when we skip prediction, as with most z-curves for example
expected_at = lagged_times[0] + dt
if expected_at < time.time() + 45:
expected_at = time.time() + dt
else:
expected_at = time.time() + self.DELAYS[-1]
dt = self.DELAYS[-1]
return expected_at, dt
def sample(self, lagged_values, lagged_times=None, name=None, delay=None):
""" Fit an echo-state machine to predict the mean, then use standard scattering for sampling """
dt = approx_mode(np.diff(lagged_times))
k = max(1, int(delay / dt))
num_lagged = len(lagged_values)
inputs = np.ones(num_lagged) # Exogenous inputs
outputs = np.array(list(reversed(lagged_values)))
esn = ESN(n_inputs=1, n_outputs=1, **self.ecn_params)
esn.fit(inputs=inputs, outputs=outputs)
future = [v[0] for v in esn.predict(np.ones(k))]
center = future[-1]
offset = center - lagged_values[0]
scenarios = differenced_bootstrap(lagged=lagged_values,
decay=0.01,
num=self.num_predictions)
samples = [offset + s for s in scenarios]
return samples