🚗 GetAround — Delay Analysis & Pricing Prediction

---

📌 Project Overview

GetAround is a peer-to-peer car rental platform where late vehicle returns can disrupt subsequent bookings, leading to customer dissatisfaction and cancellations.

This project tackles two key product challenges:

• Operational optimisation — analysing checkout delays and simulating buffer thresholds
• Pricing optimisation — deploying a Machine Learning model via a production API

---

🔗 Live Applications

Service	Description	URL
📊 Delay Dashboard	Product analytics & delay simulation	https://huggingface.co/spaces/Dreipfelt/getaround-dashboard
💰 Pricing Demo	UI for real-time price prediction	https://huggingface.co/spaces/Dreipfelt/Getaround-Pricing
🔌 API	FastAPI prediction service	https://dreipfelt-getaround-api.hf.space
📄 API Docs	Interactive documentation	https://dreipfelt-getaround-api.hf.space/docs
💻 GitHub	Source code repository	https://github.com/Data-Science-Designer-and-Developer/Project_GetAround

---

🎯 Business Objectives

Delay Management

• Measure late returns
• Simulate threshold strategies
• Optimise trade-off between blocked rentals and solved issues

Pricing Optimisation

• Predict rental price from vehicle features
• Serve model via API
• Enable real-time decision support

---

🤖 Machine Learning API

Property	Value
Algorithm	XGBoost Regressor
Target	rental_price_per_day (€)
RMSE	16.60
MAE	10.50
R²	0.738
CV RMSE	16.86
CV RMSE std	1.27
Number of features	13

Features

model_key, mileage, engine_power, fuel, paint_color, car_type,
private_parking_available, has_gps, has_air_conditioning,
automatic_car, has_getaround_connect, has_speed_regulator, winter_tires

---

🔌 API Endpoint

POST /predict

curl -X POST "https://dreipfelt-getaround-api.hf.space/predict" \
-H "Content-Type: application/json" \
-d '{
  "input": [[
    "Citroën",
    50000,
    120,
    "diesel",
    "black",
    "sedan",
    1,
    1,
    1,
    0,
    1,
    1,
    0
  ]]
}'

Response:

{
  "prediction": [124.52]
}

---

🗂️ Repository Structure

Project_GetAround/  
│  
├── api/                          # FastAPI application (model serving)  
│   ├── app.py                    # API endpoints (/predict)  
│   ├── Dockerfile                # HF deployment config  
│   ├── pipeline.pkl              # Trained ML pipeline  
│   ├── feature_names.json        # Input feature order  
│   ├── model_metrics.json        # Model performance metrics  
│   └── requirements.txt  
│  
├── delay_dashboard/              # Streamlit app (delay analysis)  
│   ├── app.py
│   └── requirements.txt
│  
├── pricing_demo/                 # Streamlit app (price prediction UI)  
│   ├── app.py  
│   └── requirements.txt  
│  
├── notebooks/                    # Data exploration & model training  
│   ├── 01_EDA_delays.ipynb  
│   └── 02_ML_pricing.ipynb  
│  
├── .gitignore  
├── requirements-dev.txt  
└── README.md  

---

🛠️ Tech Stack

Category	Tools
Language	Python 3.10
Dashboard	Streamlit, Plotly
API	FastAPI, Uvicorn
Machine Learning	Scikit-learn, XGBoost
Deployment	Hugging Face Spaces
Version Control	Git, GitHub

---

📅 Project Timeline

Stage	Description	Estimated Duration
1. Data Exploration	EDA on delays and pricing	~3h
2. Business Analysis	Threshold simulations, revenue impact, visualisations	~3h
3. Dashboard	Streamlit development and deployment	~3h
4. Machine Learning	Feature engineering, XGBoost training and evaluation	~5h
5. FastAPI API	Development of the /predict endpoint, Dockerfile	~3h
6. HF Spaces Deployment	Configuration, testing, production release	~2h
Total		~19h

⚙️ Local Setup

git clone https://github.com/Data-Science-Designer-and-Developer/Project_GetAround.git
cd Project_GetAround

Run API

cd api
pip install -r requirements.txt
uvicorn app:app --reload

Run dashboards

cd delay_dashboard
streamlit run app.py

cd pricing_demo
streamlit run app.py

---

📈 Business Recommendation

1. Delay Management Strategy

The analysis highlights a clear trade-off between operational efficiency and customer experience:

• Increasing the minimum buffer between rentals reduces conflicts
• However, it also increases the number of blocked bookings (lost revenue)

👉 Recommendation:

• Set a default buffer around 60–90 minutes

• This range provides a strong balance:

  • significant reduction in conflicts
  • limited impact on booking volume

👉 Scope recommendation:

• Apply stricter thresholds to Connect vehicles
• Keep more flexibility for Mobile check-ins

Why: Connect vehicles show more predictable behaviour, making stricter rules more effective with lower operational risk.

---

2. Pricing Strategy

The Machine Learning model enables data-driven pricing decisions:

• Price is strongly influenced by:

  • vehicle type
  • engine power
  • equipment (GPS, AC, etc.)

• Significant price variability exists for similar vehicles

👉 Recommendation:

• Use the model as a price recommendation tool for owners

• Integrate it directly into the platform to:

  • standardise pricing
  • reduce underpricing / overpricing
  • improve marketplace consistency

---

3. Product Impact

By combining both solutions, GetAround can:

• Reduce customer friction and cancellations
• Improve fleet utilisation
• Increase owner revenue through better pricing
• Enable data-driven product decisions

---

4. Next Steps

To maximise impact, the following improvements are recommended:

• Introduce dynamic buffer thresholds (adaptive to context: location, demand, vehicle type)
• Monitor A/B test performance of different delay strategies
• Continuously retrain the pricing model with fresh data
• Integrate both tools into a unified internal product dashboard

---

5. Key Takeaway

This project demonstrates how combining product analytics and machine learning can directly support strategic decisions and improve both user experience and business performance.

---

🎯 Executive Summary

GetAround is losing customers due to late returns between consecutive bookings. This project addresses two concrete product questions:

What minimum buffer time should be enforced between two bookings to reduce conflicts without significantly impacting revenue? → Analysis of 2017 data shows that a 60 to 90-minute buffer significantly reduces problematic cases while limiting the impact on revenue. How can owners optimise their listed prices? → An XGBoost model (R² = 0.74) predicts the optimal daily price based on vehicle characteristics, with a median error of approximately €10 per day.

Both tools are deployed in production and accessible via the links below.

---

🔒 GDPR Compliance

This project is carried out in an educational context using datasets provided by GetAround to Jedha Bootcamp.

Nature of the data: The datasets used (get_around_delay_analysis.xlsx, get_around_pricing_project.csv) are pseudonymised: no names, email addresses, phone numbers, or direct identifiers of drivers or owners are included. Rental identifiers (rental_id, car_id) are technical keys with no link to identifiable individuals.

Legal basis for processing: Internal analysis for service improvement purposes (legitimate interest — Art. 6(1)(f) GDPR). The data is not collected as part of this project but reused for analytical purposes.

Storage: The deployed API does not store any data transmitted through /predict requests. No personal data is retained on the server side.

Data subject rights: GetAround users have rights of access, rectification, and erasure directly with GetAround (the data controller). This project does not act as a data controller.

---

👤 Author

Frédéric Tellier
CDSD Candidate — Data Scientist
Jedha Bootcamp

LinkedIn: https://www.linkedin.com/in/frédéric-tellier-8a9170283/
Portfolio: https://github.com/Dreipfelt

CDSD Certification Project — Bloc 5 - Data Science Designer & Developer (RNCP35288)