jane-street-market-prediction

Coworking of UNIST SDMLAB.

Youngin Kwon, Yeongho Lee, MD KHALEQUZZAMAN CHOWDHURY SAYEM, MUBARRAT CHOWDHURY

🚩 Competition info

🏷️ ​Name

Jane Street Market Prediction on Kaggle

🔍 Objective

Make a Prediction for trading action using trading opportunities

⏱️ Timeline

Nov 24, 2020 - Feb 22 2021 (UTC)

Actual participation, Jan 13 2021

🗓️ ​Overall Schedule

• 1st week: Understanding about Competition with EDA, Implementation using baseline code per solutions

  • Youngin: LSTM
  • Yeongho: XGBoost
  • Sayem: Comprehensive Presentation
  • Mubarrat: LGBM

• 2nd week: [Enhancing Performance] Implementation of baseline code individually like above

• 3rd week: Best score solution Implementation

• 4th week: [Enhancing Performance] Best score solution Implementation

  • Youngin: Resnet1dcnn, ResnetLinear, EmbedNN

• 5th week: Parameter tuning for Best score solution

  Resnet1dcnn

  layers	hidden_layers	f_act	dropout	optimizer	learning_rate	weight_decay
  [5,5,5]	[512,128]	ReLU	0.2	Adam	1e-3	1e-5

  ResnetLinear

  elected	Mean AUC	Mean Utility	Fold1-AUC	Fold1-Utility	Fold2-AUC	Fold2-Utility	Fold3-AUC	Fold3-Utility	hidden-layer	n_layers	decreasing	f_act	dropout	embed_dim	optimizer	learning_rate	weight_decay
  False	0.5309	6162.2528	0.5299	6433.2630	0.5313	5603.7364	0.5315	6449.7590	512	3	True	LeakyReLU	0.34213845887711536	10	Adam	0.0009437366580626903	1.0288953711004482e-08
  True	0.5308	6433.8982	0.5301	6495.6230	0.5319	6123.3656	0.5305	6682.7062	256	2	False	SiLU	0.49627361377205387	0	Adam	1.3352033297894747e-05	8.62843672831598e-08

  EmbedNN

  Selected	Mean AUC	Mean Utility	Fold1-AUC	Fold1-Utility	Fold2-AUC	Fold2-Utility	Fold3-AUC	Fold3-Utility	hidden-layer	n_layers	decreasing	f_act	dropout	embed_dim	optimizer	learning_rate	weight_decay
  False	0.5311	6058.2003	0.5323	6253.5662	0.5307	5494.9595	0.5303	6426.0753	256	3	False	SiLU	0.23308511537027937	10	Adam	0.000663767918321238	2.6504094565959894e-07
  True	0.5326	6055.6161	0.5322	5822.0162	0.5320	5811.6806	0.5338	6533.1516	256	4	True	SiLU	0.17971171427796284	5	Adam	2.9521544108896628e-05	5.679142529741758e-05

📢 ​Repository Rule

👷 Structure

+-- input
|   +-- data/competitions/jane-street-market-prediction (competition dataset directory)
|   +-- kaggle.json (need for using kaggle API installing competition dataset)
+-- ipynb_notebooks (member's local code directory) 
|   +-- youngin
|   +-- yeongho
|   +-- sayem
|   +-- mubarrat
+-- output
|   +-- model (save pretrained model for inference)
|   +-- result (save .csv result file after inference)
+-- imgs (imgs for repository)
+-- README.md
+-- data_loader.py (methods for loading data)
+-- data_utils.py (methods for data preprocessing / model training / model inference)
+-- models.py (classes for Neural Net models)
+-- best_params.py (classes for storing best parameters)
+-- cv.py (classes for splitting Time-Series Aware Cross Validation method)
+-- train.py (python script for executing training process)
+-- inference.py (python script for executing inference process)

🌴 ​Branches

Branch is tool of github for cooperation.

- for handling admission to master branch, efficient version management (need to find how to use)

• jerry: Youngin-Kwon
• ho: Yeongho-Lee
• sayem: MD KHALEQUZZAMAN CHOWDHURY SAYEM
• mubart: MUBARRAT CHOWDHURY

🔧 Installed Packages

name	version
datatable	0.11.1
kaggle	1.5.10
numba	0.51.2
numpy	1.20.1
pandas	1.2.2
pytorch	1.7.1
scikit-learn	0.24.1
zipfile38	0.0.3

🔌 ​Implementation

train.py

• options

  • --model_type [Resnet1dcnn, ResnetLinear, EmbedNN]

    : define Neural Net model for training

  • --cv_type [SGCV, GTCV, random**(Resnet1dcnn only)**]

    : define Cross Validation Type for dataset

  • --selection [1,2] (ResnetLinear, EmbedNN only)

    : define best parameter

inference.py

• options

  • --model_type [Resnet1dcnn, ResnetLinear, EmbedNN]

    : define Neural Net model for training

  • --submit_type [csv, package**(Linux only)**]

    : define submission type

  • --cv_type [SGCV, GTCV, random**(Resnet1dcnn only)**]

    : define Cross Validation Type for dataset

  • --selection [1,2] (ResnetLinear, EmbedNN only)

    : define best parameter

  • --pretrained [True, False]

    : define use pretrained weight for inference

Examples

• train Resnet1dcnn model

  python train.py --model_type Resnet1dcnn --cv_type random

• train ResnetLinear model

  python train.py --model_type ResnetLinear --cv_type SGCV --selection 1

• (no trained .pth file) inference Resnet1dcnn model

  python inference.py --model_type Resnet1dcnn --submit_type csv --cv_type random --pretraind False

• (There is trained .pth file) inference Resnet1dcnn model

  python inference.py --model_type Resnet1dcnn --submit_type csv --cv_type random --pretraind True