SoE JNU EN502 Monsoon 2025 Machine Learning Project

Commodity-Prediction-Challenge

Project Summary: Approach & Solution Workflow

Data Loading

• Implemented an automatic loader that detects whether the notebook is running on Kaggle or locally and loads the dataset accordingly.

Exploratory Data Analysis

• Performed targeted EDA including exchange-group distributions, pair-level target analysis, time-series inspection of selected assets, and spread visualizations across multiple target pairs.

Competition Metric

• Implemented the custom RankCorr-Sharpe evaluation function as required by the competition.

Dataset Preparation

• Applied preprocessing, feature engineering (lags, rolling statistics, etc.), and created train–test splits for modeling.

Model Development

• Trained four models:

  • RandomForest Regressor

  • CatBoost Regressor

  • Single multi-output LightGBM

  • Multiple single-output LightGBM models

Model Training & Evaluation

• The RandomForest baseline achieved:
  • TRAIN RankCorr-Sharpe: 6.8863
  • TEST RankCorr-Sharpe: 1.9417
• Additional cross-validation was performed using the single-output LightGBM models.

Prediction Pipeline

• Implemented the final predict() function for seamless Kaggle inference.

Conclusion

• Although the model performance was modest, the project provided strong learning on financial time-series modeling, custom metrics, and multi-output ML pipelines. Future improvements will continue refining features and model architectures.