Datasets

We prepared a few boilerplate code examples to get you up and running with your first submission and to give you ideas on how to improve your score:

basic example – make your first submission incorporating macroeconomic features;

advanced example 1 – hyper-parameter tuning for a higher score;

advanced example 2 – NaN imputation with gradient boosting and mean-target-encoding of categorical variables.

Feel free to use it as a starting point and tinker on it to get better results!

Train Dataset

The train dataset contains financial data points for 19,733 publicly traded companies based on their quarterly and annual financial reports. This dataset is compiled using the 5 latest quarterly reports and 4 latest annual reports, and reflects financial components extracted from their corresponding balance sheet and income statements.

Test Dataset

The test dataset contains 24,119 companies with their corresponding features in the same format as the train dataset.

Macro Dataset

Two separate files (to match with train and test datasets) are provided which feature 1450 columns with different macroeconomical indicators collected considering the dates of each company's quarterly reports. For example a feature 'Federal Government Current Expenditures_Q_0_min_180_days'reflects the minimum value of Federal Government Current Expenditures for 180 days prior to Q_0. Columns names reflect the actual indicators (not obfuscated). Data in columns is normalized. companyId is a key to match the rows in the macro_train.csv and macro_forward_looking.csvwith X_train.csvand X_forward_looking.csv accordingly. Try to use different indicators to enrich your training/testing data.

The features ordering in train/test data is as follows:

Columns starting with Q_($n$) (where $n$ is the number of the quarter) contain the companies' quarterly reported financial components.

Columns starting with Y_($n$) (where $n$ is the number of the annual report) contain the companies' annually reported financial components.

Other columns represent metadata per each company.

Columns starting with:

• Q_0 are only present in targets_train.csv: contain financial components for the latest (closest to today) quarter

• Q_1 are a part of X_train.csv and X_forward_looking.csv: contain financial components of the quarter which went before Q_0

• Q_4 are a part of X_train.csv and X_forward_looking.csv: contain financial components of the furthest reported quarter, 4 quarters before Q_0

• Y_0 are a part of X_train.csv and X_forward_looking.csv: contain financial components from the latest annual report

• Y_3 are a part of X_train.csv and X_forward_looking.csv: contain financial components from the furthest annual report, 3 years before Y_0

Each quarter and each year (except for Q_0) contains 141 financial components, please refer to the data_dictionary.txt for details.

There are 17 targets (train_targets.csv) which represent the latest financial data points for each company. Participants need to train model(s) which will map the historical financial performance of the companies (X_train.csv) to their latest financial indicators.

Files

• X_train.csv - training features

• targets_train.csv - training targets

• X_forward_looking.csv - testing features

• sample_submission.csv - a sample submission file in the correct format

• macro_train.csv - macroeconomical data for companies in the X_train.csv

• macro_forward_looking.csv - macroeconomical data for companies in the X_forward_looking.csv

• data_dictionary.txt - detailed data points description

Data archive can be downloaded here

Please refer to data_dictionary.txt for detailed columns description.