Part of the The Bayesian Investor — A Primer for the CIO series.

Use Case 2 — “Which Factors Actually Matter?”: IPO Pricing and Feature Selection

What it does. When a company goes public, there are dozens of candidate explanations for how much its shares will pop on day one. This tool takes all the candidates, lets the model itself discover which few genuinely drive the outcome, and then groups the deals into peer clusters discovered in the data rather than taken from an industry label.

The method. (Bayesian horseshoe regression for feature selection, then a stochastic block model for data-driven peer clustering.)

The analogy — the needle-finder. Imagine a workshop table covered in hundreds of dials, and you’re told only a handful actually do anything; the rest are decorative. The Bayesian tool is a mechanic who starts with the dials all turned to “off,” then lets the evidence earn each dial being turned on. A dial only lights up if the data strongly insists it matters. You’re left with a short, trustworthy list of the dials that actually work — and a confidence level for each. A traditional approach would either guess which dials matter in advance (introducing human bias) or turn everything on (flooding the analysis with noise).

The analogy — the talent scout vs. the HR form. Traditional peer groups come from a form: “what industry is this company in?” But the companies that actually behave alike may not share an industry sticker. The talent scout watches how people actually perform and groups them by what they genuinely have in common — a young company with an aggressive investor behaves like another young company in a totally different industry. This tool finds those real peer groups in the data, so “comparable companies” means companies that actually compare.

Why it matters for investing. Deciding which IPOs to participate in is a one-off, high-uncertainty decision — precisely the territory where belief-probability shines. Knowing which few drivers actually matter (and how confident we are), and grouping deals by genuine similarity, gives your team an edge in allocation and underwriting that a sector-label analysis can’t provide.

Why it beats the traditional approach. Traditional regression keeps every candidate factor in the model, so noise factors muddy the signal and confidence statements are weak. The Bayesian version automatically selects the few real factors, attaches an honest probability to each, and builds peer groups from actual behavior rather than arbitrary labels — all without an analyst pre-committing to a model.