What Makes a Peer? Valuation-Anchored Similarity in Private Markets (arXiv (Cornell University))

Summary

• The study analyzes a global private-market universe of approximately 270,000 companies, including more than 53,000 firms with post-money valuations.p.1
• The final dataset features 27 variables, of which 23 are categorical, capturing complex taxonomic characteristics of private firms.p.3
• Model calibration utilizes an 80:20 train-test split to evaluate predictive and similarity accuracy out-of-sample.p.5
• Deal type serves as the dominant feature shaping valuation predictions, contributing a mean absolute SHAP value of 1.31.p.6
• The top ten represented countries account for approximately 90% of the out-of-sample evaluation dataset.p.6

Key insights

1. Supervised similarity learning defines peer firms based on shared valuation drivers rather than static geometric or semantic distance metrics.p.1
2. Tree-importance-weighted leaf-node co-occurrences in GBDTs accommodate missing data and non-linear interactions without requiring standard imputation.p.4
3. Multiplicative calibration factors of 1.74 for Series transactions and 5.06 for Unspecified transactions correct log-retransformation bias when converting back to linear values.p.6
4. Conformal prediction intervals achieve 93% empirical coverage at a 95% targeted confidence level across 53.8k test instances.p.6
5. Prediction error is concentrated at extreme size deciles, with RMSE rising to 2.35 in Q1 and 2.05 in Q10 compared to 0.94 in Q8.p.6
6. Downstream k-NN valuation tasks using the valuation-anchored similarity metric yield lower MAE and RMSE than cosine, Gower, or Euclidean distance metrics.p.7