smape-metric

This project builds a machine learning model for Zyfra to predict gold recovery from ore. It aims to optimize production efficiency and remove unprofitable parameters by modeling rougher and final recovery values. Models are evaluated using sMAPE to select the best-performing solution.

A project finished 02-22-2026 as part of the TripleTen Data Science program using real-world data and mimicking real-world project requirements. Task was to build a machine learning model capable of predicting gold recovery rates from gold ore using over 80 parameters recorded in the factory, totaling >1.5 million entries. Final sMAPE was 0.286.

Predict gold recovery at rougher and final stages using a multi-output regression model, evaluated with sMAPE and improved via hyperparameter tuning.

Optimización de la eficiencia en la purificación de oro utilizando modelos de Regresión y métricas personalizadas (sMAPE).