🚀 Predicting Startup Success with Crunchbase Data

🚀 Best Model Accuracy (Random Forest with SMOTE): 87.9%

📌 Project Overview

As someone transitioning into data science with an interest in the banking and finance sector, I sought to take on a project that combines both business relevance and technical depth. This project focuses on predicting startup success using Crunchbase data. I approached it from three key angles:

• Regression to understand funding patterns
• Classification to predict startup outcomes
• Clustering to discover natural groupings among startups

The goal was to determine the factors that increase the likelihood of a startup's success — measured by IPOs, acquisitions, or a combination of both — while incorporating both machine learning techniques and business intelligence.

---

🧰 Code and Resources Used

• Python Version: 3.11
• IDE: Jupyter Notebook
• Libraries: pandas, numpy, matplotlib, seaborn, statsmodels, scikit-learn, imbalanced-learn
• Dataset: An excerpt of a dataset released to the public by Crunchbase

---

🧹 Data Cleaning

• Filtered out companies younger than 3 years or older than 7 years to reduce noise (from 31,000+ to ~9,400 observations)
• Created a new status label:
  • Success = Acquired, IPO, or both
  • Failure = Closed or no exit
• Engineered number_degrees from MBA, PhD, MS, and Other
• Dropped columns with high missing values (e.g., acquired_companies, products_number)
• Handled missing values selectively
• Removed funding outliers (above 99th percentile)

---

📊 Exploratory Data Analysis (EDA)

EDA focused on understanding class balance and numerical relationships:

• Status Value Counts: Success vs Failure
• Log-Transformed Funding Distribution
• Correlation Heatmap of numeric variables

---

📈 Regression Modeling

🔹 Multiple Linear Regression (Statsmodels)

• Dependent variable: average_funded
• Significant predictors: average_participants, number_degrees, ipo
• offices and is_acquired were not statistically significant

---

🔹 Linear Regression (Scikit-learn)

• Repeated regression using Scikit-learn’s LinearRegression
• Log-transformed average_funded for normality
• Coefficients confirmed the importance of average_participants and ipo

---

🧪 Classification Modeling

🔸 Logistic Regression (Baseline – No SMOTE)

• Initially trained on imbalanced classes
• High overall accuracy but poor recall for “Success”
• Ineffective at identifying successful startups

---

🔸 Logistic Regression (With SMOTE)

• Used SMOTE to balance the dataset
• Improved recall and precision for the “Success” class
• More reliable at flagging high-potential startups

---

🌲 Random Forest Classification

• Used SMOTE-balanced dataset
• Boosted accuracy from 70.7% to 87%
• Strong performance across all evaluation metrics

---

🛠️ Hyperparameter Tuning

• Tuned n_estimators, max_depth, and min_samples_split using GridSearchCV
• Slight performance improvement
• Feature importance rankings remained stable

---

📚 Clustering Analysis (KMeans)

For a final unsupervised learning step, I applied KMeans Clustering to group startups based on:

• category_code (encoded)
• average_funded
• average_participants

Steps:

• Scaled all features using StandardScaler
• Used the elbow method to determine 3 optimal clusters
• Visualized results using 2D scatter plots

Key Insights:

• Startups in similar sectors tend to attract similar funding amounts
• Higher average_participants is associated with higher funding clusters
• These clusters help spot high-potential startups independent of IPO or acquisition status

---

💡 Final Thoughts

This end-to-end machine learning project gave me the chance to:

• Work with a real-world business dataset
• Build interpretable and high-performing models
• Apply data cleaning, feature engineering, and EDA effectively
• Handle class imbalance with SMOTE
• Improve model performance through hyperparameter tuning
• Add unsupervised clustering to surface hidden patterns

---