7th

JPX Kaggle Challenge 8th Place Solution Code

Note: JPX data has to be placed in a directory called /input/jpx-tokyo-stock-exchange-prediction/. It is the same structure as on the Kaggle competition site.

Hardware

Besides the Kaggle resources, we used a Intel E5-2620v4 CPU and a Nvidia GTX 1080 TI as GPU to test our approaches.

OS

All experiments were executed under Debian buster Version 10

Setup Conda Environment:

conda env create -f environment.yml
conda activate jpx

All results on the public testset from the supplementary files can be reproduced with the following bash commands executed from the root directory.

mkdir ./results # create dir to save results
python evaluate_models.py -m lgbmhierarch -p ./results -dp input/jpx-tokyo-stock-exchange-prediction/

Training

The model can be trained using the evaluation.ipynbnotebook or by creating a new file and executing the following code:

from models import LGBMHierarchModel

train, test = data_pipeline(dir_path="./input/jpx-tokyo-stock-exchange-prediction")
seed = 69
model = LGBMHierarchModel(seed=seed)
model.train(train=train, use_params=True)

Testing

Initialize the model as described above and use the predict method of the model to measure performance on test data.

performance = model.predict(test, {"rmse":(mean_squared_error, {"squared":False})})