def test_shrinkage_coef(self):
"""Dynamic vs at construction."""
n_samples = 3
n_channels = 4
n_assets = 5
x = torch.rand((n_samples, n_channels, n_assets)) * 100
layer_1 = CovarianceMatrix(sqrt=False,
shrinkage_strategy='diagonal',
shrinkage_coef=0.3)
layer_2 = CovarianceMatrix(sqrt=False,
shrinkage_strategy='diagonal',
shrinkage_coef=None)
assert torch.allclose(
layer_1(x), layer_2(x, 0.3 * torch.ones(n_samples, dtype=x.dtype)))
x_ = torch.rand((n_channels, n_assets)) * 100
x_stacked = torch.stack([x_, x_, x_], dim=0)
res_1 = layer_1(x_stacked)
res_2 = layer_2(x_stacked, torch.tensor([0.2, 0.3, 0.6]))
assert torch.allclose(res_1[0], res_1[1])
assert torch.allclose(res_1[1], res_1[2])
assert torch.allclose(res_1[2], res_1[0])
assert not torch.allclose(res_2[0], res_2[1])
assert not torch.allclose(res_2[1], res_2[2])
assert not torch.allclose(res_2[2], res_2[0])
def test_n_parameters(self, sqrt):
layer = CovarianceMatrix()
n_parameters = sum(p.numel() for p in layer.parameters()
if p.requires_grad)
assert n_parameters == 0
def test_wrong_construction(self):
with pytest.raises(ValueError):
CovarianceMatrix(shrinkage_strategy='fake')
with pytest.raises(ValueError):
layer = CovarianceMatrix(shrinkage_coef=None)
layer(torch.ones(1, 5, 2))
def test_basic(self, Xy_dummy, sqrt, shrinkage_strategy):
X, _, _, _ = Xy_dummy
n_samples, n_channels, lookback, n_assets = X.shape
X_ = X.mean(dim=1)
if n_channels == 1:
with pytest.raises(ZeroDivisionError):
CovarianceMatrix(sqrt, shrinkage_strategy=shrinkage_strategy)(X_)
else:
out = CovarianceMatrix(sqrt, shrinkage_strategy=shrinkage_strategy)(X_)
assert out.shape == (n_samples, n_assets, n_assets)
def test_sqrt_works(self):
n_samples = 3
n_channels = 4
n_assets = 5
x = torch.rand((n_samples, n_channels, n_assets)) * 100
cov = CovarianceMatrix(sqrt=False)(x)
cov_sqrt = CovarianceMatrix(sqrt=True)(x)
assert (n_samples, n_assets, n_assets) == cov.shape == cov_sqrt.shape
for i in range(n_samples):
assert torch.allclose(cov[i], cov_sqrt[i] @ cov_sqrt[i], atol=1e-2)
def __init__(
self,
n_external,
n_assets,
n_channels=5,
hidden_size=32,
max_weight=0.15,
force_symmetric=True,
p=0.2,
):
super().__init__()
self.force_symmetric = force_symmetric
self.matrix = torch.nn.Parameter(torch.eye(n_assets),
requires_grad=True)
self.exp_returns = torch.nn.Parameter(torch.zeros(n_assets),
requires_grad=True)
self.norm_layer = torch.nn.InstanceNorm2d(n_channels, affine=True)
self.collapse_external = AverageCollapse()
self.transform_layer = RNN(n_channels, hidden_size=hidden_size)
self.dropout_layer = torch.nn.Dropout(p=p)
self.dropout_layer2 = torch.nn.Dropout(p=p)
self.time_collapse_layer = AttentionCollapse(n_channels=hidden_size)
self.conv1 = Conv(n_input_channels=hidden_size,
n_output_channels=1,
method="1D")
# self.conv2 = Conv(n_input_channels=3, n_output_channels=1, method="1D")
self.linear_transform = torch.nn.Linear(n_external, n_assets)
self.linear_2 = torch.nn.Linear(n_external, n_assets)
self.covariance_layer = CovarianceMatrix(sqrt=False,
shrinkage_strategy="diagonal")
self.gamma_sqrt = torch.nn.Parameter(torch.ones(1), requires_grad=True)
self.alpha = torch.nn.Parameter(torch.ones(1), requires_grad=True)
self.preliminar_weights_layer = NumericalMarkowitz(
n_assets, max_weight=max_weight)
def stupid_compute(w, y):
"""Straightforward implementation with list comprehensions."""
from deepdow.layers import CovarianceMatrix
n_samples, n_assets = w.shape
covar = CovarianceMatrix(sqrt=False)(y[:, 0, ...]) # returns_channel=0
var = torch.cat([(w[[i]] @ covar[i]) @ w[[i]].permute(1, 0)
for i in range(n_samples)],
dim=0)
vol = torch.sqrt(var)
lhs = vol / n_assets
rhs = torch.cat([(1 / vol[i]) * w[[i]] * (w[[i]] @ covar[i])
for i in range(n_samples)],
dim=0)
res = torch.tensor([((lhs[i] - rhs[i])**2).sum()
for i in range(n_samples)])
return res