BMW Car Sales Analysis & Price Prediction

Temporal Validation of Market Value

Min Set Khant (Solo)

Introduction

Project: BMW Car Sales — Data Analysis & Price Prediction

• Goal: Build a robust, temporally validated model to predict BMW car prices using multi-regional sales data.
• Key Questions Addressed:
  1. What are the key features influencing BMW car prices?
  2. Can a Machine Learning model accurately predict prices?
  3. How stable are these predictions over time (Temporal Validation)?

Data Overview & Setup

Dataset: BMW Worldwide Sales (2010–2024)

• Source: Kaggle (Multi-Regional Sales Data)
• Status: Data has been cleaned, imputed, and One-Hot Encoded.
• Target Variable: Price_USD (Log-transformed for modeling).

	Model	Year	Region	Color	Fuel_Type	Transmission	Engine_Size_L	Mileage_KM	Price_USD	Sales_Volume	Sales_Classification
0	5 Series	2016	Asia	Red	Petrol	Manual	3.5	151748	98740	8300	High
1	i8	2013	North America	Red	Hybrid	Automatic	1.6	121671	79219	3428	Low
2	5 Series	2022	North America	Blue	Petrol	Automatic	4.5	10991	113265	6994	Low
3	X3	2024	Middle East	Blue	Petrol	Automatic	1.7	27255	60971	4047	Low
4	7 Series	2020	South America	Black	Diesel	Manual	2.1	122131	49898	3080	Low
5	5 Series	2017	Middle East	Silver	Diesel	Manual	1.9	171362	42926	1232	Low
6	i8	2022	Europe	White	Diesel	Manual	1.8	196741	55064	7949	High
7	M5	2014	Asia	Black	Diesel	Automatic	1.6	121156	102778	632	Low
8	X3	2016	South America	White	Diesel	Automatic	1.7	48073	116482	8944	High
9	i8	2019	Europe	White	Electric	Manual	3.0	35700	96257	4411	Low

EDA: Market Trends & Segmentation

Regional Performance and Price Trends

Sales Trend Over Time	Top Models by Regional Sales Volume
Observation: The average price shows a clear, non-linear upward trend over time, suggesting market inflation. This justifies the Temporal Validation strategy.	Observation: Model popularity is highly segmented by region. This confirms that the Region variable is a critical feature for effective price segmentation.

Detail: The analysis utilized over 50,000 records of multi-regional sales. Regional features were processed using One-Hot Encoding to prevent models from assigning arbitrary ordinal relationships to categories.

Global Market Demand

Global Market Demand: Top 10 Models (2010–2024)

Volume by Model (2010–2024)
Insight:Globally, sales volume is often dominated by core sedan and SUV series (e.g., 3-Series, 5-Series, X3, X5), reflecting high overall market liquidity.

Impact:Models with high global volume typically exhibit more stable pricing due to consistent demand, which reduces prediction volatility. EDA Detail:This volume data was analyzed before the temporal split to understand the underlying market foundation across all years. |

EDA: Correlation Analysis

Price Relationships

Correlation Matrix	Price Distribution by Fuel Type
Observation: Log-Price shows a strong inverse linear relationship with Car Age (depreciation) and a strong positive correlation with Engine Size (performance/class).	Observation: Significant differences in the median price across fuel types confirm this categorical feature has strong predictive power and is not independent of the target.

Detail: Two key features were engineered for this analysis: Car Age (calculated as Current_Year - Model_Year) and Price per KM (calculated as Price_USD / Mileage_KM), both showing significant predictive value.

Modeling Approach

Temporal Validation Strategy (The Key Differentiator)

• Goal: Test model stability and real-world predictive ability.
• Split: - Train Set: All data from 2010–2023.
  • Test Set: Data from Year 2024 (The “future” market).
• Model Chosen: Random Forest Regressor (highest predictive power).
• Evaluation: MAE and RMSE are converted back to USD for business interpretability.

Random Forest Performance

Model Accuracy (Reporting in USD)

Metric	Value	Interpretation
R-squared (Log Price)	0.9952	Variance explained by the model on the transformed price.
Root Mean Squared Error (RMSE)	$2,292.13	Average prediction error, penalizing large errors.
Mean Absolute Error (MAE)	$531.13	Most Interpretable: On average, the model’s price prediction is off by this amount in USD.

Interpretation: The model successfully predicted 2024 prices with an average absolute error of approximately $531.13, proving its temporal stability.

Key Price Drivers: Feature Importance

Top 5 Predictors of BMW Price

The Random Forest model identified these key drivers based on Gini Impurity Reduction:

1. Price per KM (0.682) - Represents the calculated efficiency/condition ratio; its dominance suggests it captures the residual value most effectively.
2. Mileage (KM) (0.318) - Primary measure of vehicle use and wear, directly driving the rate of depreciation.
3. Engine Size (L) (0.000018)
4. Car Age (0.000015)
5. Transmission: Manual (0.000004)

Full Feature Importance

Technical Detail: The model was chosen for its ability to calculate Feature Importance via Mean Decrease in Impurity (Gini). The trained model was serialized and saved as rf_best_model.joblib for future deployment.

Conclusion: Summary of Findings

Answering the Key Questions

• Top Drivers: Price is overwhelmingly driven by engineered features: Price per KM and Mileage (KM).
• Model Success: The Random Forest model demonstrates strong temporal validation, accurately predicting 2024 market prices with an MAE of $531.13.
• Market Insight: Model popularity and sales volume are highly specific to Region.

Business Recommendations

Actionable Insights for Strategy

1. Pricing Strategy: Use the model’s feature importance to set pricing guidelines. The low MAE ($531.13) provides high confidence for automated pricing models.
2. Inventory Optimization: Focus inventory based on the regional popularity derived from the EDA.
3. Future Validation: Re-run the temporal validation annually to ensure the model’s stability and adapt to market shifts.

Thank You

INFO-523 Final Project

• Notebooks: 01_Data_Preparation_and_EDA.ipynb, 02-modeling.ipynb
• Output: Trained model saved as rf_best_model.joblib.
• Data: Cleaned data saved as bmw_cleaned.csv and bmw_modeling_ready.csv.