def test_degenerate_col(self):
batch, height, width, degenerate = 6, 7, 8, 4
x = get_degenerate_sample(batch, height, width, degenerate, 69)
inputs = [x[..., 0] for x in tf.split(x, width, axis=-1)]
normaliser = preprocessing.MeanVarianceNormaliser()
outputs = normaliser.fit_transform(inputs)
for step, op in enumerate(outputs):
assert_near(tf.reduce_mean(op, 0), 0.)
expected_var = 0. if step == degenerate else 1.
assert_near(tf.math.reduce_variance(op, 0), expected_var)
assert_near(normaliser.inverse_transform(outputs), inputs)
def test_low_variance(self):
batch, height, width = 2**20, 3, 2
x = get_low_variance_sample(batch, height, width, 69)
inputs = [x[..., 0] for x in tf.split(x, width, axis=-1)]
normaliser = preprocessing.MeanVarianceNormaliser()
outputs = normaliser.fit_transform(inputs)
for op in outputs:
tf.debugging.assert_type(op, FLOAT_DTYPE)
# test with FLOAT_DTYPE
mean, variance = tf.nn.moments(op, 0)
assert_near(mean, 0.)
assert_near(variance, 1., atol=1e-2)
# test with tf.float64
mean, variance = tf.nn.moments(tf.cast(op, tf.float64), 0)
assert_near(mean, 0., atol=1e-8)
assert_near(variance, 1., atol=1e-8)
assert_near(normaliser.inverse_transform(outputs), inputs)
risk_neutral=True,
learning_rate=1e-1)
driver.verbose = 2
risklevels = [0.05, 0.5, 0.95] if not cost else [0.95]
for alpha in risklevels:
driver.add_testcase(f"deep network {alpha}",
hedge_models.NeuralHedge(
timesteps=timesteps,
instrument_dim=book.instrument_dim,
internal_dim=0,
num_layers=4,
num_units=5,
activation=tf.keras.activations.softplus),
risk_measure=hedge_models.ExpectedShortfall(alpha),
normaliser=preprocessing.MeanVarianceNormaliser(),
feature_function="log_martingale",
price_type="arbitrage")
if driver.cost is not None or not driver.risk_neutral:
driver.add_liability_free(
hedge_models.LinearFeatureHedge(
timesteps=timesteps,
instrument_dim=book.instrument_dim,
mappings=[approximators.IdentityFeatureMap] \
* (1 + (driver.cost is not None))),
risk_measure=hedge_models.ExpectedShortfall(alpha),
normaliser=preprocessing.MeanVarianceNormaliser(),
feature_function="log_martingale")
train_size, test_size = int(2**18), int(2**18)