Versioning

Logs the master dataset changes

Master 1.0.3 - 2023/04/14

Train - Test split

• X_train_full and y_train_full files were added for download. The last 3 months' data and resolved targets are included. Unresolved targets are set as NaNs.

Master 1.0.2 - 2023/04/07

Features

• One feature from the gordon strategy was removed due to a minor fix.

Master 1.0.1 - 2023/03/31

Features

• Some features were removed from the master dataset. The features names changed since they are named by their order in their respective strategies.

Master 1.0 - 2023/03/24

Features

• The features processing has changed and the features are not orthogonalized to risk factors anymore.

Targets

• The features changed from specific compounded returns over the target's span to raw compounded returns over the respective span.

Master 0.1 - 2023/03/17

Features

• Feature names range from Feature_1 to Feature_n. They are prefixed with their respective strategy.

• Features are orthogonal to risk factors.

• The quantization of the features is done in a gaussian way by moon's cross-sections.

Targets

• Computed as the cumulative product of the specific return of each asset on different time horizons, with a 2 days lag:

  • target_w: 7 days

  • target_r: 30 days

  • target_g: 60 days

  • target_b: 90 days

• The targets are quantized in a fat-tailed way by moon cross-sections.

Train - Test split

• The dataset is split into three parts:

  • X_train contains the features up to the last moon - 90 days.

  • y_train contains the targets up to the last moon - 90 days.

  • X_test contains the features of the last 90 days.

• Embargo: the 90 days size of X_test data you must predict represents the part of the data on which the targets are not fully computed - target_b being a 90 days target.

• Moons are continuous between the train set and the test set - First moon of the test set starts at max moon of the train set + 1.

Example submission

The example_submission file is built with a linear regression built with sklearn.

import pandas as pd
from sklearn.linear_model import LinearRegression


X_train = pd.read_parquet('X_train.parquet')
y_train = pd.read_parquet('y_train.parquet')
X_test = pd.read_parquet('X_test.parquet')

train_set = pd.merge(X_train, y_train, on=['id', 'Moons'])
train_set.dropna(inplace=True)

targets_cols = [c for c in train_set.columns if 'target' in c]
features_cols = [c for c in train_set.columns if 'Feature' in c]
    
model = LinearRegression().fit(train_set.loc[:, features_cols], train_set.loc[:, targets_cols])
predictions = pd.DataFrame(model.predict(X_test.loc[:, features_cols]), columns=targets_cols)

example_submission = pd.concat([X_test.iloc[:, :2], predictions], axis=1)

example_submission.to_csv('example_submission.csv', index=False)
example_submission.to_parquet('example_submission.parquet')

Submission

Submissions file must respect the following rules to be valid:

• The column names must be the same as those in the example_submission file.

• Values must be in the 0-1 range.

• NaNs are not accepted.

• The number of moons must be the same as the ones in the example_submission file.

• The ids in each moon must be the same as in the example_submission file.

• There mustn't be any constant column in any column in any moon.