Analysis of Italian MICHELIN restaurants

Project Overview

This project aims to predict the Return on Assets (ROA) for 2022 using various machine learning techniques.

• Project Overview
• Data Description
• Project Structure
• Feature Engineering
• Modeling Approaches
• Results
• License

Data Description

The dataset used in this analysis includes various financial metrics and other relevant features. Key variables include:

• ROA (%): Return on Assets from the year 2020 to 2022.
• Indice_Liquidità: Index of liquidity from 2020 to 2022.
• Indice_Finanziario: Index of financial independence from 2020 to 2022.
• Follower_Restaurant and Follower_Chef: Instagram follower for restaurant and chef from 2018 to 2022. It contains a lot of missing data so in the end we didn't use this data for our predictive analysis
• Type of Structure: Hotel, Catering, Guest house, Farmhouse
• Star: Number of star from 1 to 3
• N. of employees: to understand the size of the company
• Social Capital
• Dummies variables for regions.

Feature Engineering

We created two new variables to summarize the effect of the two index previously calculated that are more statistically significant.

• Delta_Liquidità_2022: Change in liquidity in 2022.
• Delta_Finanziario_2022: Change in financial independence in 2022.

Modeling Approaches

Several models were implemented and compared:

OLS Regression: The baseline model used for prediction.

Ridge Regression with Spline Transformation: To capture non-linearities.

Polynomial Features: To capture non-linear relationships between features.

Simple Decision Tree: Trying a tree based method to see how it performs.

Random Forest: Improving performance of tree base method using RF.

Results

Empirical Results: R-squared is the metric used to evaluate the performance of each model

Visual Results: Visualizing the predicted y versus the actual y provided a clear view of how well the regression line corresponds with our data.

Explanable AI: SHAP was used to interpret the OLS model, understand feature importance and more importantly how much and how each variable influence the final result.

License

I'd like to thanks my friend Simone Staiano to manually scrape data all over the web to provide me this Dataset, all the credits goes to him.