Startup Risk Factor Analysis Using Chaid Decision Tree in Entrepreneurial Funding Platforms

Entrepreneurial funding platforms spend limited due diligence time on many startups, with a lot of the due diligence parameters being categorical or ordinal as opposed to continuous (e.g. tiers of founder experience, ranges of number of employees, etc.) and with the investment committees needing to segment according to logic they can easily explain in a clear and comprehensible fashion rather than a nebulous score function. Structurally, CHAID (Kass 1980) is an appropriate algorithm to address this requirement: CHAID is a chi-squared splitting algorithm that is different from binary decision trees in that it allows for multi-way splits in which statistically indistinguishable categories are combined, yet genuinely distinct categories are separated. This paper applies CHAID to outcome data from 8,400 funded startups on an entrepreneurial funding platform (Success: 4,704, 56.0%; Failure: 3,696, 44.0%, observed over a 36-month tracking window), using eight categorical risk factors spanning founder background, financial discipline, and market structure. The resulting tree's first split, on founder prior startup experience$\left(\chi^{2}=284.6, \mathrm{p}<0.001\right)$, separates success rates of 42.4% (no prior experience), 58.8% (one prior venture), and 74.4% (two or more); compounding the no-experience branch with high burn rate (>10% of raised capital monthly) and high customer concentration (>50% revenue from one customer) identifies an 8.4%-success critical-risk segment, while two-or-more-prior-venture founders in high-revenue-multiple sectors reach 82.4% success-a 74.0 percentage point spread between the tree's best and worst terminal segments. CHAID achieves 74.8% predictive accuracy$(\text{AUC}=0.798)$using only 18 terminal nodes, outperforming logistic regression (68.4%; +9.4%) and a binary CART tree (72.4%; +3.3%) while trailing a 500-tree Random Forest (78.4%; −4.8%) that offers no comparable business-interpretable structure. Deployed to triage due-diligence effort, CHAID-derived risk tiers reduced mean review time by 38.1% (8.4 to 5.2 hours per application) while validating prospectively (68.4% subsequent failure rate among funded critical-tier applications versus 16.8% among low-tier), though survivorship bias, reverse-causality ambiguity in the tree's correlational structure, a 49.3% disparate-impact pattern in scrutiny-tier assignment by founder background, and substantial regime-dependence in risk-factor strength across the 2022 venture funding contraction each qualify the model's deployment scope.

Hali tarjima qilinmagan