In our last post, we looked at accuracy on motor vehicle accident cases using Median Absolute Percentage Error (MdAPE). MdAPE shows how close predictions are in a typical case, expressed as Median Accuracy — the right central-tendency metric for a long-tail damages distribution.
Here, we apply the same approach to premises liability (PL) cases. The PL fold of the model is trained on 72,000 PL verdicts and settlements — meaningfully smaller than the MVA fold’s 198,000, which has implications for the confidence-band width but, as the per-tier numbers below show, not for the headline accuracy.
Median accuracy across premises liability cases
The per-tier breakdown:
| Case value range | Median accuracy |
|---|---|
| $0 – $60,000 | 90.0% |
| $60,000 – $210,000 | 89.9% |
| $210,000 – $400,000 | 88.5% |
| $400,000 – $840,000 | 86.9% |
| $840,000 – $2,000,000 | 92.4% |
| $2,000,000 – $5,000,000 | 93.1% |
What this shows
Accuracy is consistent across the range
Across most premises cases, median accuracy stays in a tight band between roughly 87% and 90%. That consistency means the model produces values close to the actual outcome at every case-value tier — not just at the tier the model was optimized for.
For an attorney making case-selection decisions across a mixed PL caseload, the consistency matters as much as the headline accuracy. A model that performs well at one tier and poorly at another would force the attorney to remember which tier they’re in; a model that performs consistently across the range can be trusted to produce a defensible number without that mental overhead.
Accuracy improves where it matters most
At higher case values, median accuracy moves into the low 90s — exceeding 93% at the top end. These are the cases where precision matters most in negotiation, and they are also the most consistent. The pattern is structural: higher-value cases tend to involve clearer severity classifications and denser comparable-verdict data, which the model’s stratified jurisdiction folds exploit.
The 86.9% trough at the $400K–$840K range is worth flagging honestly. That tier sits at the boundary between “moderate severity” and “severe injury” classifications in our schema, and the variance widens at the boundary. We’ve documented the boundary case explicitly in the methodology page; the recalibration policy means this tier is a high-priority target for the next quarterly model refresh.
What this means in practice
Evaluate cases faster
When the model produces a defensible value at the intake call, the case-selection decision moves from gut to evidence in 30 seconds. The defended number arrives at intake — before the retainer is signed and the firm’s hours are committed.
A defensible frame for negotiation
A tight accuracy band gives plaintiff counsel a defensible position when setting demand-letter values, evaluating opposing offers, and deciding how to proceed. The carrier’s reserve number stops being the only anchor in the room.
Why PL bands run wider than MVA bands
A note for attorneys who work both case types. The PL accuracy numbers above are headline-comparable to the MVA accuracy numbers in the previous post — both run in the high 80s to low 90s. But the confidence bands around each PL prediction run modestly wider than the bands around MVA predictions at the same severity tier.
The reason: data density. The PL training set carries 72,000 cases versus the MVA fold’s 198,000. Confidence bands narrow with density — the model’s quantile-regression band layer reflects the data thickness. Same headline accuracy figure; wider band per prediction. Both are honest, and the band is the methodology.
For the per-jurisdiction PL breakdown — comparative-fault rules, no-fault MVA effects on PL economics, state-by-state median settlement values — the state-by-state hub carries the per-state pages.
Final thought
Across both MVA and premises liability, Predict delivers consistently high median accuracy with performance improving where the stakes are highest. By calibrating against a held-out test set with stratified jurisdiction folds, the model produces settlement values that hold up across the kinds of cases plaintiff PI attorneys actually see — not just the canonical mid-tier rear-end MVA in dense-data states.
To run the model on a real PL case from your intake pipeline, the free calculator is open. The full Predict model — with comparable-verdict citations, demand-letter-ready valuation blocks, and quarterly recalibration — runs inside the 14-day free trial. No credit card at signup.