/ projects · 01
Predicting startup acquisition vs. closure
Given only what a startup looks like on paper (sector, funding history, geography), can you tell whether it gets acquired or quietly shuts down? I trained Random Forest and SVM classifiers on Crunchbase data to find out. Try the distilled version below.
/ live demo · distilled from the real model
Sector odds are the paper's real acquisition rates (Enterprise Software 81% → Clean Tech 39%); the funding effect is centred between the real medians.
The data
The raw export holds 54,294 companies across 39 columns. Most are still operating, outcomes unknown, so the modelling set is the 6,295 companies with a resolved ending: 3,692 acquired, 2,603 closed. Cleaning had its moments: funding amounts arrived with mangled comma placement ("17,50,000"), and 15% of rows had no usable funding total at all. After cleaning, 5,375 companies and 25 features remained, split 59.9% acquired / 40.1% closed.
The models
An untuned 100-tree Random Forest memorised the training set (99% train vs 71% test accuracy). Capping depth at 10 closed the gap to 80/74; a grid search over depth, split size and tree count confirmed that was the right cap. The RBF-kernel SVM, trained on scaled features, managed 66%. Five-fold cross-validation put the forest's ROC-AUC at 0.77 ± 0.02.
What actually matters
- Total funding dominates: importance 0.31, exactly double the next feature. Acquired startups had a median raise of $8.4M vs $1.0M for closed ones (Mann-Whitney p < 10⁻²¹⁵).
- Venture backing comes second at 0.15, then market (0.09) and the number of funding rounds (0.08).
- Sector spread is wide: Enterprise Software startups were acquired 81% of the time; Clean Technology, 39%.
About the demo
The widget above is a small logistic model built from these findings: real sector rates, real funding medians, honest spread. The full Random Forest lives in a Jupyter notebook; request the paper →