def test_errors(self, dataloader_dummy, network_dummy):
with pytest.raises(TypeError):
generate_metrics_table({'bm_1': 'WRONG'}, dataloader_dummy, {'metric': MeanReturns()})
with pytest.raises(TypeError):
generate_metrics_table({'bm_1': network_dummy}, 'FAKE', {'metric': MeanReturns()})
with pytest.raises(TypeError):
generate_metrics_table({'bm_1': network_dummy}, dataloader_dummy, {'metric': 'FAKE'})
def test_wrong_construction_1(self, dataloader_dummy):
"""Wrong positional arguments."""
with pytest.raises(TypeError):
Run('this_is_fake', MeanReturns(), dataloader_dummy)
with pytest.raises(TypeError):
Run(DummyNet(), 'this_is_fake', dataloader_dummy)
with pytest.raises(TypeError):
Run(DummyNet(), MeanReturns(), 'this_is_fake')
def test_wrong_construction_2(self, dataloader_dummy):
"""Wrong keyword arguments."""
with pytest.raises(TypeError):
Run(DummyNet(), MeanReturns(), dataloader_dummy, metrics='this_is_fake')
with pytest.raises(TypeError):
Run(DummyNet(), MeanReturns(), dataloader_dummy, metrics={'a': 'this_is_fake'})
with pytest.raises(ValueError):
Run(DummyNet(), MeanReturns(), dataloader_dummy, metrics={'loss': MeanReturns()})
with pytest.raises(TypeError):
Run(DummyNet(), MeanReturns(), dataloader_dummy, val_dataloaders='this_is_fake')
with pytest.raises(TypeError):
Run(DummyNet(), MeanReturns(), dataloader_dummy, val_dataloaders={'val': 'this_is_fake'})
with pytest.raises(TypeError):
Run(DummyNet(), MeanReturns(), dataloader_dummy, benchmarks='this_is_fake')
with pytest.raises(TypeError):
Run(DummyNet(), MeanReturns(), dataloader_dummy, benchmarks={'uniform': 'this_is_fake'})
with pytest.raises(ValueError):
Run(DummyNet(), MeanReturns(), dataloader_dummy, benchmarks={'main': OneOverN()})
def test_basic(self, dataloader_dummy, network_dummy):
metrics_table = generate_metrics_table({'bm_1': network_dummy},
dataloader_dummy,
{'rets': MeanReturns()})
assert isinstance(metrics_table, pd.DataFrame)
assert len(metrics_table) == len(dataloader_dummy.dataset)
assert {'metric', 'value', 'benchmark', 'timestamp'} == set(metrics_table.columns.to_list())
def test_launch_interrupt(self, dataloader_dummy, monkeypatch):
network = DummyNet(n_channels=dataloader_dummy.dataset.X.shape[1])
loss = MeanReturns()
class TempCallback(Callback):
def on_train_begin(self, metadata):
raise KeyboardInterrupt()
monkeypatch.setattr('time.sleep', lambda x: None)
run = Run(network, loss, dataloader_dummy, callbacks=[TempCallback()])
run.launch(n_epochs=1)
def run_dummy(dataloader_dummy, network_dummy, Xy_dummy):
""""""
X_batch, y_batch, timestamps, asset_names = Xy_dummy
device = X_batch.device
dtype = X_batch.dtype
return Run(network_dummy,
MeanReturns(),
dataloader_dummy,
val_dataloaders={'val': dataloader_dummy},
benchmarks={'bm': OneOverN()},
device=device,
dtype=dtype)
def test_attributes_after_construction(self, dataloader_dummy, additional_kwargs):
network = DummyNet()
loss = MeanReturns()
kwargs = {}
if additional_kwargs:
kwargs.update({'metrics': {'std': StandardDeviation()},
'val_dataloaders': {'val': dataloader_dummy},
'benchmarks': {'whatever': OneOverN()}})
run = Run(network, loss, dataloader_dummy, **kwargs)
assert network is run.network
assert loss is run.loss
assert dataloader_dummy is run.train_dataloader
assert isinstance(run.metrics, dict)
assert isinstance(run.val_dataloaders, dict)
def test_launch(self, dataloader_dummy):
network = DummyNet(n_channels=dataloader_dummy.dataset.X.shape[1])
loss = MeanReturns()
run = Run(network, loss, dataloader_dummy)
run.launch(n_epochs=1)
dataloader = RigidDataLoader(dataset,
indices=list(range(5000)),
batch_size=batch_size,
lookback=lookback)
val_dataloaders = {
'train':
dataloader,
'val':
RigidDataLoader(dataset,
indices=list(range(5020, 9800)),
batch_size=batch_size,
lookback=lookback)
}
run = Run(network,
100 * MeanReturns(),
dataloader,
val_dataloaders=val_dataloaders,
metrics={'sqweights': SquaredWeights()},
benchmarks={
'1overN': OneOverN(),
'VAR': VARTrue(process),
'Random': Random(),
'InverseVol': InverseVolatility()
},
optimizer=torch.optim.Adam(network.parameters(), amsgrad=True),
callbacks=[EarlyStoppingCallback('val', 'loss')])
history = run.launch(40)
fig, ax = plt.subplots(1, 1)
X = np.stack(X_list, axis=0)[:, None, ...]
y = np.stack(y_list, axis=0)[:, None, ...]
dataset = InRAMDataset(X, y, asset_names=returns.columns)
dataloader = RigidDataLoader(dataset, batch_size=batch_size)
# %%
# The main feature of :code:`deepdow` is that it only cares about the final allocation that minimizes
# some function of **empirical** portfolio returns. Unlike with the sample estimators in the previous
# sections, one does not need to explicitly model the dynamics of the market and find allocation via
# the two step procedure. Below we define empirical counterparts of the convex optimization objectives
# (losses). Note that all losses in :code:`deepdow` have the *the lower the better* logic.
all_losses = {
'minvar': StandardDeviation()**2,
'maxret': MeanReturns(),
'maxutil': MeanReturns() + gamma * StandardDeviation()**2
}
# %%
# Training and evaluation
# ^^^^^^^^^^^^^^^^^^^^^^^
# Now it is time to train!
deep_portfolios_c = {}
deep_portfolios_u = {}
for mode in ['u', 'c']:
for loss_name, loss in all_losses.items():
network = Net(n_assets, max_weight=max_weight if mode == 'c' else 1.)
run = Run(
# dropout it is essential that we set the mode correctly based on what we are trying to do.
network = network.train(
) # it is the default, however, just to make the distinction clear
# %%
# Training
# ^^^^^^^^
# It is now time to define our loss. Let's say we want to achieve multiple objectives at the same
# time. We want to minimize the drawdowns, maximize the mean returns and also maximize the Sharpe
# ratio. All of these losses are implemented in :code:`deepdow.losses`. To avoid confusion, they
# are always implemented in a way that **the lower the value of the loss the better**. To combine
# multiple objectives we can simply sum all of the individual losses. Similarly, if we want to
# assign more importance to one of them we can achieve this by multiplying by a constant. To learn
# more see :ref:`losses`.
loss = MaximumDrawdown() + 2 * MeanReturns() + SharpeRatio()
# %%
# Note that by default all the losses assume that we input logarithmic returns
# (:code:`input_type='log'`) and that they are in the 0th channel (:code:`returns_channel=0`).
# %%
# We now have all the ingredients ready for training of the neural network. :code:`deepdow` implements
# a simple wrapper :code:`Run` that implements the training loop and a minimal callback
# framework. For further information see :ref:`experiments`.
run = Run(network,
loss,
dataloader_train,
val_dataloaders={'test': dataloader_test},
optimizer=torch.optim.Adam(network.parameters(), amsgrad=True),