- Pools are priced on averages. Returns are produced by loans. The variance tax is what sits in between.
- Two pools with the same weighted-average LTV do not perform the same, because loss severity is not linear in LTV. The average sits where nothing happens and reports nothing about the borrowers that do the damage.
- The error only ever runs one way. A loan repays par at best, so good borrowers cannot compensate for bad ones. Averaging flatters the pool, and the overstatement grows with the dispersion it discarded.
- Where the pool is tranched, the same principle applies again and changes sign across the capital stack — so the resolution matters more, not less.
- None of this is recoverable downstream. It is destroyed at the point of averaging, before pricing begins.
Pools are priced on averages. Returns are produced by loans. Everything that follows sits in that gap.
MinVaris underwrites loan by loan — 130 million individual loans across 2,000 deals — because the information that determines realised performance is destroyed by the act of averaging, before pricing begins. This note sets out what gets destroyed, and why no amount of downstream modelling recovers it.
What an average destroys
Take two mortgage pools with the same weighted-average LTV of 75%. In the first, borrowers cluster near that number. In the second, half sit around 60% and half around 90%. Identical headline statistic, identical presentation in a marketing deck.
They will not perform the same, because loss severity is not linear in LTV. Below roughly 80%, a fall in house prices is absorbed by borrower equity and the lender is made whole. Above it, the lender starts taking the loss, and the further above, the faster it accumulates. The dispersed pool has real weight in that region. The tight pool has none at all. On a simple model of house price shocks and forced-sale costs, that difference alone carries the dispersed pool to something like 40–50% more expected loss — on identical reported statistics.
Why the error only runs one way
The obvious objection is that the 60% half compensates for the 90% half. It does not. A loan repays principal and coupon at best. There is no outcome in which a low-LTV borrower pays back more than contracted to offset one who defaults. The upside is capped; the downside is not.
That asymmetry is why averaging is not a neutral simplification. Good loans cannot rescue bad ones, so the error runs in a single direction: pricing off the average flatters the pool, and the overstatement grows with exactly the dispersion the average threw away.
The same holds for debt service ratios, seasoning, income verification and payment history. Each has a non-linear relationship to loss. Each is routinely reported as one number.
Where the pool is tranched, it happens twice
In securitised form the principle applies a second time, and does something less intuitive. A tranche absorbs losses between its attachment and detachment points, which gives it two corners bending opposite ways. The attachment corner costs the holder as losses spread out, because more of the distribution reaches into the band. The detachment corner pays, because losses beyond it belong to someone else.
Plot that sensitivity across the stack and it is zero at both ends, peaking at expected losses. Everything above is flattered by central-case pricing, which is why super-senior paper looked close to free in 2006. First loss is penalised by it, charged for severity it can never absorb once written off.
What only the tape shows
The shape of that distribution is set by how correlated the defaults are, and correlation is not a number to be taken from a table. It is clustering, and clustering is visible only borrower by borrower.
Shared regional labour markets. Concentration in a single broker channel. Cohorts underwritten to one standard at one point in the cycle. Servicers with materially different cure rates on identical arrears buckets. Jurisdictions where enforcement runs years longer, turning a modest difference in defaults into a large one in losses.
None of it survives a weighted average. All of it sits in the loan tape, where it goes unpriced — and unlike a public-market signal, it is not competed away.
The losses come from how they differ.