A settlement estimate delivered as a single dollar figure invites a decision calibrated to a precision the underlying data cannot support. The number looks authoritative because it is specific, and specificity reads as confidence even when the process that produced it collapsed a wide range of plausible outcomes into one point chosen for convenience rather than for accuracy. An institution that prices a reserve, a facility advance, or a portfolio position off that single figure has adopted the figure's false precision as its own, and the error does not announce itself until the matter resolves somewhere else on the range the point estimate never disclosed.
Settlement and award values do not cluster around a mean the way many financial variables do. The distribution is bounded below by zero or by a nominal nuisance figure, and it carries a long right tail produced by the minority of matters that resolve well above the median, a large verdict, an aggregated claim, a bellwether result that resets a defendant's settlement posture across an entire docket. A distribution with this shape has a mean pulled upward by the tail and a median that sits meaningfully below it, and reporting only one of those two numbers, without the shape connecting them, misrepresents what a typical matter in the population actually resolves to.
What a properly built model produces is a labeled distribution, not a claim about what one specific matter will realize in dollars. The distribution is constructed from a corpus of comparable matters that actually reached resolution, tagged by the procedural and factual characteristics that made them comparable, with every input distinguishing what a seller stated about a matter from what the corpus independently corroborates. This labeling discipline matters because a distribution that quietly blends seller-asserted values with verified resolution values is not measuring the same thing at every point along its range, and an institution pricing off it inherits that inconsistency without being told it exists.
Two matters can share an identical expected settlement value and represent entirely different underwriting risk once the shape of their distributions is examined. One matter's value clusters tightly around its median, with a narrow band reflecting a deep comparable population and a well-defined procedural posture. Another matter's value splits between a modest majority outcome and a small-probability, large-magnitude tail, with almost no mass in between. Pricing both matters off their shared expected value treats them as the same position. They are not, and the difference is precisely the information a distribution carries and a point estimate discards.
Institutional underwriting should draw its number from a stated percentile of the distribution, not from its mean, and the choice of percentile should reflect the capital's actual risk tolerance rather than a default convention applied without examination. A conservative facility reserving against downside exposure should price closer to a lower percentile, accepting a wider gap to the mean in exchange for a smaller probability of being wrong on the low side. A position seeking upside participation can reasonably price closer to the mean, but should disclose that it is doing so and should disclose the tail it is choosing to accept alongside that choice.
The width of a settlement band is itself a piece of underwriting information, not a modeling shortcoming to be minimized at all costs. A band that narrows as more comparable, resolved matters accumulate behind it reflects a model doing its job, converging toward a defensible range as evidence accumulates. A band that stays wide despite a large comparable population reflects genuine heterogeneity in how that category of matter resolves, and narrowing it artificially, by reporting a false point estimate instead of the honest band, does not remove the underlying uncertainty. It only hides it from the desk that has to price against it.
None of this licenses a claim that the model predicts what any single matter will actually realize in dollars, and that distinction is not a technicality. A labeled distribution states what a population of comparable, resolved matters has historically realized, conditioned on stated inputs; it does not state that this particular pending matter will realize a value inside that range, only that the range is the best evidence available for where it is likely to fall. Return metrics computed downstream, an IRR, a MOIC, are computed from the distribution plus the capital's stated terms, structure, and cost, not asserted as direct model outputs, and a vendor or a desk that blurs that computation into the model's claim is overstating what the model actually knows.
The labeling discipline extends to how comparables are selected in the first place, not only to how their values are reported. A comparable population assembled loosely, on broad case-type similarity alone, will produce a wider and less useful band than one assembled against the specific procedural and factual characteristics that actually drive settlement value in that category of matter, jurisdiction, defendant type, injury or damages class, procedural posture at the time of valuation. An institution should ask not just how many comparable matters sit behind a distribution but what specific characteristics defined comparability, because a loosely defined comparable population produces a band that looks scientific while carrying much of the same imprecision as an outright guess.
A related discipline applies to how a distribution is updated as a matter progresses rather than left static from the moment of intake. A settlement distribution estimated at filing should narrow, or at minimum shift, as procedural milestones accumulate additional information about the matter's likely trajectory, and a vendor whose reported distribution never changes between intake and resolution is either not re-scoring the position at all or is reporting a number that was never genuinely sensitive to new evidence in the first place.
An institutional buyer evaluating a settlement estimate should ask for the band, the comparable population size behind it, and the percentile the reported figure represents, not for a single number delivered with confidence. A vendor that can produce all three has priced a distribution honestly. A vendor that can produce only a dollar figure has produced a guess formatted to look like a measurement, and the two should never be priced the same.