bike-sales-profit-model

🚲 Bike Sales Profit Prediction

🎥 Project Video Presentation : https://www.loom.com/share/01d0945ce32c43e1b765cf901f92e325

📦 Project Overview

This project focuses on predicting delivery days and classifying delivery speed for e-commerce orders using a cleaned and engineered dataset. The project includes two main tasks:

Regression: Predicting the exact number of delivery days.

Classification: Categorizing deliveries into Fast, Medium, or Slow groups.

📊 The Dataset

Source: Olist Brazilian E-Commerce Public Dataset (after preprocessing).

Size: ~100,000 rows after cleaning.

Target Variables:

delivery_days (Regression)

delivery_class (Classification: 0 = Fast, 1 = Medium, 2 = Slow)

Key Features Used

customer_state – destination state (logistics relevance)

product_weight_g – product weight

freight_value – shipping cost

price – product price

cluster_label – group assigned via K-Means clustering

🛠️ Feature Engineering & Cleaning

To improve model performance, several steps were applied:

1. Data Cleaning

Removed negative or unrealistic delivery values.

Removed extreme outliers in profit and delivery_days using IQR.

1. Location Encoding

Applied One-Hot Encoding to the customer_state column.

1. K-Means Clustering

Created 4 clusters based on weight, price, and freight cost.

Added the cluster label as a new feature for both tasks.

🤖 Model Performance Part 1: Regres

sion — Predicting Delivery Days

We compared several regression models on engineered features.

Insight: Feature engineering (especially clustering + location encoding) clearly improved predictive accuracy. ==========

🧠 Classification Performance (Fast / Medium / Slow)

To classify delivery speed into Fast, Medium, and Slow, three different classification models were trained:

Logistic Regression

Random Forest Classifier

Gradient Boosting Classifier

Each model was evaluated using Accuracy and F1-macro, ensuring a balanced view of performance across all three classes.

🏆 Best Model

The Random Forest Classifier achieved perfect performance across all evaluation metrics and was selected as the final winning model.

It was exported and uploaded as:winning_classifier.pkl

📁 Repository Structure

This repository contains:

Assignment_2.ipynb — full notebook with EDA, preprocessing, modeling, and evaluation.

winning_model.pkl — trained Gradient Boosting regression model.

winning_classifier.pkl — trained XGBoost classification model.

README.md — project documentation (this file).