LEGARDIA
Data Analytics · Retail
Score any product from a new collection — predicted tier, confidence, and recommended action.
Live ML model — requires the FastAPI backend running locally on port 8000. Full source on GitHub ↗.
Product Details
Fill in the product details and run the model.
Methodology
Training
Logistic Regression on 1,936 real SKUs from FS24 and SS25. SMOTE balanced classes to 683 samples per tier. GridSearchCV selected C=10, L2 penalty, saga solver.
Threshold
Dead stock probability threshold tuned to 0.36 instead of the default 0.5. This achieves 89% recall on dead stock — the model catches almost every case, at the cost of some average-class false positives.
Fuzzy Matching
Color and collection names are fuzzy-matched to training vocabulary. "Wine Red" maps to "Burgundy", "rainforest" maps to "Rainforest". Below 75% similarity, the value passes through unmatched.