Inspiration
Mango faces the challenge of accurately forecasting clothing production to minimize overstock and shortages, specially focusing on shortage. Our goal was to use data science techniques to optimize inventory
What it does
Our solution predicts the optimal production quantity for each clothing item using historical sales, product features, and embedding images. It outputs production to balance the risk of excess inventory and stock outs, focusing more on avoiding shortages.
How we built it
Data Preprocessing: Cleaned and engineered features from raw sales and product data, including embeddings for image features and temporal variables.
Feature Engineering: Applied normalization, one-hot encoding for categorical variables, and dimensionality reduction for embeddings.
Modeling: Trained LightGBM and XGBoost quantile regression models with custom asymmetric loss functions to penalize shortages more heavily than excess.
Validation: Used custom metrics and cross-validation to ensure robust performance.
Submission Pipeline: Automated prediction and CSV generation for competition submission.
Challenges we ran into
Handling high-cardinality categorical variables and missing data.
Aligning train/test features after complex preprocessing.
Designing a custom loss function that reflects costs, with higher weight for shortages.
Efficiently processing large datasets with embedding features.
Accomplishments that we're proud of
Built a robust, modular pipeline for data cleaning, feature engineering, and modeling.
Successfully integrated image embeddings, numerical, categorical and temporal features.
Developed a custom asymmetric metric for Mango's business needs.
Achieved strong validation results and automated the submission process.
What we learned
The importance of careful feature engineering and preprocessing in real-world retail data.
How to design and implement custom loss functions for business objectives.
Best practices for aligning train/test data and handling categorical variables.
The value of model interpretability and robust validation.
What's next for OPTIMAL INVENTORY PREDICTION DATATHON FME 2025 MANGO
Explore deep learning models for further improvements.
Integrate external data sources (e.g., weather, trends).
Enhance interpretability for business users.
Take into account the cannibalism between similar clothes sold on the same dates. Already tried to exploit this by considering the sales of similar clothes on overlap dates, but much more could be investigated.
