Context

As a Business Intelligence Analyst at ESN Data (fictional), I was asked by Jade (the director) to support “Les Plus Beaux Logis de Paris” on two consecutive missions:

• Historical analysis: study the evolution of real estate prices in Paris (2017–2021) to identify market dynamics and key correlations.
• Predictive modeling and clustering:
  • Build a linear regression model to predict the future valuation of the company’s asset portfolio.
  • Implement a K-Means clustering algorithm to automatically classify properties (apartments vs commercial spaces), addressing a request from Louise (director at “Les Plus Beaux Logis de Paris”).

Datasets

• df_histo_immo: historical property sales in Paris from 2017 to 2021 (transactions = 26,196, fields: transaction date, sale price, street number, street name, postal code, city, property type code, property type, surface)
• df_actifs: company asset portfolio for 2022, in two versions:
  • Portfolio 1: 275 properties (154 individuals, 121 corporate)
  • Portfolio 2: 469 properties (310 individuals, 159 corporate)
• Clustering sample: property data provided by Louise’s team to validate the K-Means model

Workflow

• Environment: Jupyter Notebook (Python 3) using pandas, numpy, matplotlib, seaborn, plotly, statsmodels, scikit-learn
• EDA:
  • Type checking, counting by property type
  • Time span calculation
  • Evolution of average price per m² overall and by district (2017→2021)
  • Representativeness analysis (apartments vs commercial)
  • Pearson correlation tests: price/m² vs date, price/m² vs surface
  • Price/m² comparison between apartments and commercial units
• Linear regression:
  • Preprocessing: log transformation of target variable, one-hot encoding of postal codes, time-based train/test split (train ≤ 2020, test = 2021)
  • Model training with LinearRegression, evaluation using MAE as percentage
  • Validation goal: <10% mean error
• K-Means clustering:
  • Preparation: compute price per m², handle missing values, variable selection
  • K-Means application (k=2), cluster interpretation, mapping to categories (apartment vs commercial)
  • Evaluation: f1-score, accuracy, confusion matrix

Insights

• Market trends: Prices rose steadily before COVID, then slowed down. Each district has its own price level but follows a similar trend.
• Representativeness:
  • 2021 historical data: 92.9% apartments vs 7.1% commercial
  • Portfolio 1: 56.0% apartments vs 44.0% commercial
  • Portfolio 2: 66.1% apartments vs 33.9% commercial
• Predictions:
  • Portfolio 1 valuation: €67.5M (individual) | €111.9M (corporate)
  • Portfolio 2 valuation: €788.5M (individual) | €52,454.8M (corporate)
  • Weighted mean error: 1.08% (train), 1.11% (test)
• Clustering:
  • f1-score = 1.00; accuracy = 100%
  • Perfect classification (confusion matrix with no errors)

Business Impact

• Executives now have a solid historical analysis to understand price dynamics by district
• The regression model anticipates portfolio valuation with controlled error (<1.5%), supporting investment decisions
• Automated clustering replaces manual categorization, saving time and ensuring reliable property classification

Links

Notebook (EDA)

Presentation