← All insights

Why age, not just fund group, matters

The fund-group accuracy breakdown we publish holds every Buyout fund in the held-out set to one number, regardless of whether it was priced at year 4 or year 9 of its life. That hides something real: a 4-year-old fund has had far less time to reveal what it's actually going to return than a 9-year-old one nearing the end of its life, where most of the value has already either been distributed or written down. If accuracy quietly gets easier as funds age, an aggregate number flatters the youngest, hardest cases — which also happen to be the ones a secondary buyer prices most often, since older funds have less unrealized book left to sell.

Re-running the same test, split by age

We took the identical out-of-sample battery behind our published verdicts — calibrate on vintages through 2010, price held-out 2011-2016 Buyout funds with no look-ahead — and split the results into two cuts: funds priced at year 4-6 specifically, versus the full year 4-9 window we publish today. Same funds, same curves, same models; only the age filter changes.

Typical miss is the median absolute out-of-sample error on TVPI. Both models' error is roughly 60% larger in the year 4-6 cut than in the published all-ages aggregate — the same direction, for both an independently-fit baseline and the model we actually ship.

The range gets less reliable too, not just the point estimate

Coverage tells the same story from a different angle: how often the published P25-P75 range actually contained the realized outcome (targets 50% by construction). For Equisect Bayesian it drops from 51% across all ages to 47% in the year 4-6 slice specifically — close to target either way, but weaker exactly where it matters most. The TA baseline's band stays wide enough to still over-cover at young ages (69% vs. 74%), which is a different failure mode: not miscalibrated so much as wide enough to hide a noisier point estimate underneath it. Rank-IC — whether funds predicted to do relatively better actually did — softens too: 0.51 to 0.43 for Equisect Bayesian, 0.52 to 0.44 for TA.

Why this happens

It isn't a modeling bug — it's what limited data actually looks like. A fund at year 4-6 has typically called most of its committed capital but distributed comparatively little of it back; almost all of its eventual outcome is still sitting in unrealized NAV and marks, not in cash that's already happened. There's simply less realized signal to calibrate a price against, for every model, not just ours. The published aggregate isn't wrong — it's an honest average across an age range where the hard part (early) and the easy part (late) blend together.

What to do with it

If you're pricing a Buyout stake at year 4-6 specifically, treat the P25-P75 range as doing more of the work than the single fair-price number — even more than our standing house rule already says for Buyout in general. It's the same reason a fund only a few quarters past its first capital call gets the widest bands the platform produces: less realized history earns a wider honest range, not a falsely precise one.

Research and software, not investment advice.