Where our pricing accuracy is strongest — and where to lean harder on the range
A single out-of-sample verdict per model hides real variation underneath it. Some fund groups are just easier to price than others — here's the breakdown we don't hide, and what to do with it.
One number isn't the whole story
Every model we ship gets one headline verdict — survives or overfit — published on the landing page and re-run live against every price. That number is honest, but it's an average across every fund in the validation set. A model can clear its overfitting test overall while still being noticeably more accurate on some fund groups than others, simply because some groups have deeper history and steadier outcomes than others.
So instead of only publishing the aggregate, here's the same held-out accuracy check broken out by fund group, for Equisect Bayesian — the model with the most granular published breakdown today. (Equisect Cohort's accuracy is validated in aggregate — see why most PE pricing models fail out-of-sample — but not yet broken out by fund group the way this one is.)
The breakdown
| Fund group | Held-out funds | Typical miss | Band covers actual (target 50%) | Rank-IC |
|---|---|---|---|---|
| Growth | ~240 | 0.43× | 62% | 0.52 |
| Balanced | ~90 | 0.45× | 76% | 0.56 |
| Infrastructure | ~45 | 0.38× | 84% | 0.66 |
| Buyout | ~240 | 0.52× | 52% | 0.46 |
| Secondaries | ~30 | 0.68× | 50% | 0.78 |
| Venture Capital | ~50 | 0.84× | 34% | 0.17 |
Typical miss is the median absolute pricing error on TVPI, out-of-sample. Band covers actual is how often the published P25–P75 range actually contained the realized outcome — the model targets 50% here by construction, so a number well below that means the range was too narrow for that group specifically. Rank-IC is whether funds predicted to do relatively better actually did, regardless of the absolute price (outlier-robust, Spearman).
What this actually says
Growth, Balanced, and Infrastructure are the strongest groups: coverage at or above the 50% target, the smallest typical misses, and a real history behind each (roughly 90–240 held-out funds). Buyout sits close behind — coverage a touch under target, but a large, well-populated sample.
Venture Capital and Secondaries are the two to treat with real caution. Venture's coverage is the lowest by a wide margin — the P25–P75 range misses the actual outcome roughly two-thirds of the time it should have caught it — and the typical miss is nearly double the strongest groups'. That's not surprising: venture outcomes are famously lumpy (a small share of funds drive most of the return), and there are only around 50 held-out funds behind the number, versus roughly 240 for Growth. Secondaries sits right at the 50% coverage target but on the thinnest sample of all (around 30 funds) — treat that number as provisional until it's backed by more history.
What to do with it
The house rule already applies here — use the P25–P75 band, not the point estimate, as your honest range — but this breakdown tells you how much to lean on that advice. On Growth, Balanced, or Infrastructure, a narrow band is a genuinely informative signal. On Venture Capital, widen your own margin of error beyond what the band alone shows, and treat any single price as a starting point for diligence, not a number to trade on directly.
Research and software, not investment advice.