Walk-forward Optimization¶
Install strategy file¶
Moonshot looks for strategies inside the "moonshot" directory, so execute the following cell to "install" the strategy by moving the file to that location:
The
!sytax below lets us execute terminal commands from inside the notebook.
# make directory if doesn't exist
!mkdir -p /codeload/moonshot
!mv kitchensink_ml.py /codeload/moonshot/
Run walk-forward optimization¶
Now we can run a walk-forward optimization, which iteratively trains and tests the machine learning model over successive periods.
In this example, we retrain the model annually (train="Y").
Because we do not specify a model parameter, the default model will be used, which is a regression model using Stochastic Gradient Descent. See the usage guide for more details about the default model.
The parameter progress=True causes the moonshot service to log the walk-forward progress during the analysis. To view the progress, open a separate Terminal window and stream the logs (using cut to format the logs to fit better in the window): quantrocket flightlog stream | cut -d ' ' -f 5-
As always, it is also a good idea to stream the detailed logs (
quantrocket flightlog stream -d) to see more granular logging during the walk-forward optimization.
from quantrocket.moonshot import ml_walkforward
ml_walkforward("kitchensink-ml",
start_date="2006-12-31",
end_date="2018-12-31",
train="Y",
progress=True,
filepath_or_buffer="kitchensink_ml*")
In the terminal window, you'll see output resembling the following:
[kitchensink-ml] Walk-forward analysis progress
train test progress
start end start end status Sharpe
iteration
0 1998-12-31 2001-12-30 2001-12-31 2002-12-30 ✓ -0.66
1 2001-12-31 2002-12-30 2002-12-31 2003-12-30 ✓ 4.13
2 2002-12-31 2003-12-30 2003-12-31 2004-12-30 ✓ 0.59
3 2003-12-31 2004-12-30 2004-12-31 2005-12-30 -
4 2004-12-31 2005-12-30 2005-12-31 2006-12-30
5 2005-12-31 2006-12-30 2006-12-31 2007-12-30
6 2006-12-31 2007-12-30 2007-12-31 2008-12-30
...from moonchart import Tearsheet
Tearsheet.from_moonshot_csv("kitchensink_ml_results.csv")
Trained model¶
The trained model can be opened and inspected with joblib:
import joblib
trained_model = joblib.load("kitchensink_ml_trained_model.joblib")
print(type(trained_model))
estimator = trained_model.named_steps["estimator"]
print(estimator.coef_)
To perform live trading, you would need to update your ML strategy to point to this trained model:
class TheKitchenSinkML(MoonshotML):
...
MODEL = "/codeload/kitchensink_ml/kitchensink_ml_trained_model.joblib"
See the usage guide for more details.