def simulate_univariate_hawkes(mu, alpha, beta, run_time, seed=None):
"""
Simulates a univariate Hawkes based on the parameters.
:param mu, alpha, beta: parameters of the Hawkes process
:param run_time: End time of the simulation
:param seed: (optional) Seed for the random process.
:return: Hawkes event times
"""
# this is due to tick's implementation of Hawkes process
alpha = alpha / beta
# Hawkes simulation
n_nodes = 1 # dimension of the Hawkes process
adjacency = alpha * np.ones((n_nodes, n_nodes))
decays = beta * np.ones((n_nodes, n_nodes))
baseline = mu * np.ones(n_nodes)
hawkes_sim = SimuHawkesExpKernels(adjacency=adjacency,
decays=decays,
baseline=baseline,
verbose=False,
seed=seed)
hawkes_sim.end_time = run_time
hawkes_sim.simulate()
event_times = hawkes_sim.timestamps[0]
return event_times
def _sim_single_exposures(self):
if not self.sparse:
raise ValueError(
"'single_exposure' exposures can only be simulated"
" as sparse feature matrices")
if self.hawkes_exp_kernels is None:
np.random.seed(self.seed)
decays = .002 * np.ones((self.n_features, self.n_features))
baseline = 4 * np.random.random(self.n_features) / self.n_intervals
mult = np.random.random(self.n_features)
adjacency = mult * np.eye(self.n_features)
if self.n_correlations:
comb = list(permutations(range(self.n_features), 2))
if len(comb) > 1:
idx = itemgetter(
*np.random.choice(range(len(comb)),
size=self.n_correlations,
replace=False))
comb = idx(comb)
for i, j in comb:
adjacency[i, j] = np.random.random(1)
self._set(
'hawkes_exp_kernels',
SimuHawkesExpKernels(adjacency=adjacency,
decays=decays,
baseline=baseline,
verbose=False,
seed=self.seed))
self.hawkes_exp_kernels.adjust_spectral_radius(
.1) # TODO later: allow to change this parameter
hawkes = SimuHawkesMulti(self.hawkes_exp_kernels,
n_simulations=self.n_cases)
run_time = self.n_intervals
hawkes.end_time = [1 * run_time for _ in range(self.n_cases)]
dt = 1
self.hawkes_exp_kernels.track_intensity(dt)
hawkes.simulate()
self.hawkes_obj = hawkes
features = [[
np.min(np.floor(f)) if len(f) > 0 else -1 for f in patient_events
] for patient_events in hawkes.timestamps]
features = [
self.to_coo(feat, (run_time, self.n_features)) for feat in features
]
# Make sure patients have at least one exposure?
exposures_filter = itemgetter(
*[i for i, f in enumerate(features) if f.sum() > 0])
features = exposures_filter(features)
n_samples = len(features)
return features, n_samples
def test_hawkes_set_timestamps(self):
"""...Test simulation after some timestamps have been set manually
"""
decay = 0.5
baseline = [0.2, 0.4]
adjacency = [[0.2, 0.1], [0, 0.3]]
hawkes = SimuHawkesExpKernels(
baseline=baseline, decays=decay, adjacency=adjacency,
verbose=False, seed=1393)
# timestamps generated by a hawkes process with the same
# characteristics simulated up to time 10
original_timestamps = [
np.array([7.096244, 9.389927]),
np.array([0.436199, 0.659153, 2.622352, 3.095093,
7.189881, 8.068153, 9.240032]),
]
hawkes.track_intensity(1)
hawkes.set_timestamps(original_timestamps, 10)
hawkes.end_time = 100
hawkes.simulate()
# Intensity up to time 10 is known
first_intensities = hawkes.tracked_intensity[0][:10]
np.testing.assert_array_almost_equal(
first_intensities,
[0.2, 0.2447256, 0.27988282, 0.24845138, 0.23549475,
0.27078386, 0.26749709, 0.27473586, 0.24532959, 0.22749379]
)
# Ensure other jumps have occured afterwards
self.assertGreater(hawkes.n_total_jumps,
sum(map(len, original_timestamps)))
def SimExp(baseline, adjacency, decays, num_clusters, data):
hawkes = SimuHawkesExpKernels(adjacency=adjacency,
decays=decays,
baseline=baseline,
verbose=False)
#dt = 0.001 #millisecond granularity
#hawkes.track_intensity(dt) # turning this on will eat up memory
#need to compute and draw from the cluster length distrbution from the original data
cluster_lengths = ComputeClusterLengths(data)
multi = SimuHawkesMulti(hawkes, n_simulations=num_clusters)
multi.end_time = np.random.choice(cluster_lengths,
size=num_clusters,
replace=True)
multi.simulate()
sim_inner_timestamps = multi.timestamps
l = 0
for realisation in sim_inner_timestamps:
for series in realisation:
l += len(series)
print(f"Simulated {l} points")
return sim_inner_timestamps
def gen_event_times(self, run_time):
decays = self.parameters['decay']
adjacency = self.parameters['adjacency']
baseline = self.parameters['baseline']
hawkes_sim = SimuHawkesExpKernels(adjacency=adjacency,
decays=decays,
baseline=baseline,
verbose=False)
hawkes_sim.end_time = run_time
hawkes_sim.simulate()
# Generate Event Arrival Times
limit_order_arrivals = hawkes_sim.timestamps[1]
limit_order_arrivals = [('loa', time) for time in limit_order_arrivals]
limit_order_cancellations = hawkes_sim.timestamps[2]
limit_order_cancellations = [('loc', time)
for time in limit_order_cancellations]
market_order_arrivals = hawkes_sim.timestamps[0]
market_order_arrivals = [('moa', time)
for time in market_order_arrivals]
# Merge into single array of event times
event_times = limit_order_arrivals.copy()
event_times.extend(limit_order_cancellations)
event_times.extend(market_order_arrivals)
event_times.sort(key=lambda x: x[1])
return event_times
def setUp(self):
np.random.seed(23982)
self.n_nodes = 3
self.baseline = np.random.rand(self.n_nodes)
self.adjacency = np.random.rand(self.n_nodes, self.n_nodes) / 2
self.decays = np.random.rand(self.n_nodes, self.n_nodes)
self.adjacency[0, 0] = 0
self.adjacency[-1, -1] = 0
self.hawkes = SimuHawkesExpKernels(self.adjacency, self.decays,
baseline=self.baseline, seed=203,
verbose=False)
def fit_exp_hawkes_and_simulate(train_times, decay, end_time):
learner = HawkesExpKern(decay, verbose=True, max_iter=100000, tol=1e-10)
learner.fit(train_times)
score = learner.score()
print(f'obtained {score}\n with {decay}\n')
decay_matrix = np.full((1, 1), decay)
simulation = SimuHawkesExpKernels(learner.adjacency,
decay_matrix,
baseline=learner.baseline,
end_time=end_time)
simulation.simulate()
return learner, simulation
def get_train_data(n_nodes=3, betas=1.):
np.random.seed(130947)
baseline = np.random.rand(n_nodes)
adjacency = np.random.rand(n_nodes, n_nodes)
if isinstance(betas, (int, float)):
betas = np.ones((n_nodes, n_nodes)) * betas
sim = SimuHawkesExpKernels(adjacency=adjacency, decays=betas,
baseline=baseline, verbose=False,
seed=13487, end_time=3000)
sim.adjust_spectral_radius(0.8)
adjacency = sim.adjacency
sim.simulate()
return sim.timestamps, baseline, adjacency
def SimulateExp(baseline, adjacency, decays, time):
hawkes = SimuHawkesExpKernels(adjacency=adjacency,
decays=decays,
baseline=baseline,
verbose=False)
hawkes.end_time = time
dt = 0.001 #millisecond granularity
#hawkes.track_intensity(dt)
print(f"Starting sim")
hawkes.simulate()
timestamps = hawkes.timestamps
l = 0
for series in timestamps:
l += len(series)
print(f"Simulated {l} points")
return hawkes.timestamps
def simulate(self):
n_nodes = 1
baseline = [0.1]
adjacency = [[0.1]]
end_time = 10000
# max_jumps=1000;
a_sim = SimuHawkesExpKernels(adjacency,
decays,
baseline=baseline,
end_time=end_time,
verbose=True)
a_sim.track_intensity(0.01)
a_sim.simulate()
# print(a_sim.timestamps)
# print('Tracked intensity: ', a_sim.tracked_intensity)
with open('sample_timestamps.txt', 'w') as f:
f.write(str(list(a_sim.timestamps[0])))
def test_solver_scpg(self):
"""...Check Self-concordant proximal gradient solver for a Hawkes
model with ridge penalization
"""
beta = 3
betas = beta * np.ones((2, 2))
alphas = np.zeros((2, 2))
alphas[0, 0] = 1
alphas[0, 1] = 2
alphas[1, 1] = 3
mus = np.arange(1, 3) / 3
hawkes = SimuHawkesExpKernels(adjacency=alphas,
decays=betas,
baseline=mus,
seed=1231,
end_time=20000,
verbose=False)
hawkes.adjust_spectral_radius(0.8)
alphas = hawkes.adjacency
hawkes.simulate()
timestamps = hawkes.timestamps
model = ModelHawkesExpKernLogLik(beta).fit(timestamps)
prox = ProxL2Sq(1e-7, positive=True)
pg = SCPG(max_iter=2000, tol=1e-10, verbose=False,
step=1e-5).set_model(model).set_prox(prox)
pg.solve(np.ones(model.n_coeffs))
original_coeffs = np.hstack((mus, alphas.reshape(4)))
np.testing.assert_array_almost_equal(pg.solution,
original_coeffs,
decimal=2)
def simulate_sparse_realization(self):
"""Simulate realization in which some nodes are sometimes empty
"""
baseline = np.array([0.3, 0.001])
adjacency = np.array([[0.5, 0.8], [0., 1.3]])
sim = SimuHawkesExpKernels(adjacency=adjacency,
decays=self.decay,
baseline=baseline,
verbose=False,
seed=13487,
end_time=500)
sim.adjust_spectral_radius(0.8)
multi = SimuHawkesMulti(sim, n_simulations=100)
adjacency = sim.adjacency
multi.simulate()
# Check that some but not all realizations are empty
self.assertGreater(max(map(lambda r: len(r[1]), multi.timestamps)), 1)
self.assertEqual(min(map(lambda r: len(r[1]), multi.timestamps)), 0)
return baseline, adjacency, multi.timestamps
np.random.seed(7168)
n_nodes = 3
baselines = 0.3 * np.ones(n_nodes)
decays = 0.5 + np.random.rand(n_nodes, n_nodes)
adjacency = np.array([
[1, 1, -0.5],
[0, 1, 0],
[0, 0, 2],
], dtype=float)
adjacency /= 4
end_time = 1e5
integration_support = 5
n_realizations = 5
simu_hawkes = SimuHawkesExpKernels(
baseline=baselines, adjacency=adjacency, decays=decays,
end_time=end_time, verbose=False, seed=7168)
simu_hawkes.threshold_negative_intensity(True)
multi = SimuHawkesMulti(simu_hawkes, n_simulations=n_realizations, n_threads=-1)
multi.simulate()
nphc = HawkesCumulantMatching(integration_support, cs_ratio=.15, tol=1e-10,
step=0.3)
nphc.fit(multi.timestamps)
plot_hawkes_kernel_norms(nphc)
size=args.n_correlations,
replace=False)
comb = [comb[i] for i in idx]
for i, j in comb:
adjacency[i, j] = np.random.random()
if args.constant_decay:
decays = np.full((args.n_types, args.n_types), args.exp_decay)
else:
decays = np.random.exponential(args.exp_decay,
(args.n_types, args.n_types))
simu_hawkes = SimuHawkesExpKernels(
baseline=baseline,
adjacency=adjacency,
decays=decays,
verbose=False,
seed=args.rand_seed,
)
simu_hawkes.adjust_spectral_radius(args.adj_spectral_radius)
simu_hawkes.max_jumps = args.max_jumps
print(simu_hawkes.baseline)
print(simu_hawkes.adjacency)
print(simu_hawkes.decays)
with Timer("Simulating events"), Pool(cpu_count() // 2) as p:
timestamps = list(
starmap(
simulate_helper,
zip(
from tick.plot import plot_point_process
period_length = 100
t_values = np.linspace(0, period_length)
y_values = 0.2 * np.maximum(np.sin(t_values *
(2 * np.pi) / period_length), 0.2)
baselines = np.array(
[TimeFunction((t_values, y_values), border_type=TimeFunction.Cyclic)])
decay = 0.1
adjacency = np.array([[0.5]])
hawkes = SimuHawkesExpKernels(adjacency,
decay,
baseline=baselines,
seed=2093,
verbose=False)
hawkes.track_intensity(0.1)
hawkes.end_time = 6 * period_length
hawkes.simulate()
fig, ax = plt.subplots(1, 1, figsize=(10, 4))
plot_point_process(hawkes, ax=ax)
t_values = np.linspace(0, hawkes.end_time, 1000)
ax.plot(t_values,
hawkes.get_baseline_values(0, t_values),
label='baseline',
ls='--',
def simulate(parameters, T, adjacency, decay):
'''
parameters = (pi, beta, rho, u, delta00, delta01, delta10, delta11, mu1, mu2, mu3)
'''
pi, beta, rho, u, delta00, delta01, delta10, delta11, mu1, mu2, mu3 = parameters
baseline = np.array([mu1, mu2, mu3])
# Simulate Times
hawkes_sim = SimuHawkesExpKernels(adjacency=adjacency,
decays=decay,
baseline=baseline,
verbose=False)
hawkes_sim.end_time = T
hawkes_sim.simulate()
# Generate Event Arrival Times
limit_order_arivals = hawkes_sim.timestamps[1]
limit_order_arivals = [('loa', time) for time in limit_order_arivals]
limit_order_cancelations = hawkes_sim.timestamps[2]
limit_order_cancelations = [('loc', time)
for time in limit_order_cancelations]
market_order_arivals = hawkes_sim.timestamps[0]
market_order_arivals = [('moa', time) for time in market_order_arivals]
# Merge into single array of event times
event_times = limit_order_arivals.copy()
event_times.extend(limit_order_cancelations)
event_times.extend(market_order_arivals)
event_times.sort(key=lambda x: x[1])
#%% initialize order book
lob = OrderBook('XYZ', tick_size=1)
AAPL_LOB = pd.read_csv( # Orderbook
'../../Data/LOBSTER_SampleFile_AAPL_2012-06-21_50/AAPL_2012-06-21_34200000_37800000_orderbook_50.csv',
header=None)
init = AAPL_LOB.iloc[0]
t = 10**(-10)
dt = 10**(-10)
#Initial Orders
for i in range(0, AAPL_LOB.shape[1], 4):
if init[i + 1] > 0:
lob.submit_limitorder(-1, init[i] / 100, init[i + 1], t)
t += dt
for i in range(2, AAPL_LOB.shape[1], 4):
if init[i + 1] > 0:
lob.submit_limitorder(1, init[i] / 100, init[i + 1], t)
t += dt
midprice = [lob.mid_price()]
spread = [lob.spread()]
time = [lob.time]
for event in event_times:
# simulation
if event[0] == 'loa': # limit order arrival
order = gen_limitorder(lob, delta00, delta01, delta10, delta11,
rho, u)
ID = lob.submit_limitorder(*order, event[1])
elif event[0] == 'loc': # limit order cancellation
# get number of sell and buy orders
num_sells = sum(
[len(value) for key, value in lob.sell_side.items()])
num_buys = sum([len(value) for key, value in lob.buy_side.items()])
# cancel random order if condition met
if (num_sells > 5 and num_buys > 5):
active_orders = get_active_orders(lob)
lob.cancel_limitorder(np.random.choice(active_orders),
event[1])
else:
continue # if nothing happended dont collect data
elif event[0] == 'moa': # market order case
order = gen_market_order(lob, beta, pi)
lob.submit_marketorder(*order, event[1])
else:
raise ValueError('Invalid event type')
midprice.append(lob.mid_price())
spread.append(lob.spread())
time.append(lob.time)
Output = pd.DataFrame(index=pd.to_timedelta(time, unit='s'),
data={
'mid_price': midprice,
'spread': spread
})
sigma5 = np.std(
np.log(Output.resample('5s').first().mid_price).diff().dropna())
sigma30 = np.std(
np.log(Output.resample('30s').first().mid_price).diff().dropna())
sigma60 = np.std(
np.log(Output.resample('60s').first().mid_price).diff().dropna())
meanSpread = np.mean(Output.spread)
return (sigma5, sigma30, sigma60, meanSpread, *parameters)
def _corresponding_simu(self):
"""Create simulation object corresponding to the obtained coefficients
"""
return SimuHawkesExpKernels(adjacency=self.adjacency,
decays=self.decay,
baseline=self.baseline)
threshold = alpha / rho
ZS = soft_thres_S(Aest+US, threshold)
US = US + Aest - ZS
if __name__ == '__main__':
end_time = 50000
n_realizations = 1
decay = 3
baseline =np.ones(6) * .03
adjacency = np.zeros((6, 6))
adjacency[2:, 2:] = np.ones((4, 4)) * 0.1
adjacency[:3, :3] = np.ones((3, 3)) * 0.15
hawkes_exp_kernels = SimuHawkesExpKernels(adjacency=adjacency, decays=decay,
baseline=baseline, end_time=end_time,
verbose=False, seed=1039)
multi = SimuHawkesMulti(hawkes_exp_kernels, n_simulations=n_realizations)
multi.end_time = [(i + 1) / n_realizations * end_time for i in range(n_realizations)]
multi.simulate()
type_seq = []
time_seq = []
for i, d in enumerate(multi.timestamps[0]):
for j in d:
type_seq.append(i)
time_seq.append(j)
type_seq = np.array(type_seq)
time_seq = np.array(time_seq)
def _corresponding_simu(self):
return SimuHawkesExpKernels(adjacency=self.adjacency,
decays=self.decays,
baseline=self.baseline)