How we calculate the BaseRate

How we build it

We train a machine-learning model on millions of real negotiated prices that hospitals are federally required to publish. For any procedure, it learns the fair price from what similar hospitals actually charge — adjusting for where the hospital is (prices genuinely differ by region) and what kind of hospital it is, and anchoring to the Medicare rate for that procedure. The result is a fair-price estimate tuned to this hospital, in this place.

The range, and how sure we are

Every BaseRate comes with a typical range (for example, “$5–$6”) and a confidence level. A wider range means we’re less certain. Confidence is tied to how much real price data we have for that exact procedure at that hospital:

• High — lots of comparable prices; a tight, reliable estimate.
• Medium — a solid estimate with a moderate range.
• Low — a rarer procedure or thin data; treat the number as a rough guide, not gospel.

How accurate it is

We hold the model to an honest standard. For most procedures, the BaseRate lands within about 11% of the actual clean market price, and roughly 8 out of 10 real prices fall inside the range we show. It is deliberately conservative and gets sharper where we have more data — like any estimate, it’s a strong guide, not an appraisal.

Where the data comes from

Every price starts from a hospital’s own public price-transparency file, published under federal rule (45 CFR §180). We show the exact source file and the date it was published on each hospital page — you can open it and check the numbers yourself. Nothing here is invented.

How we clean the data

Hospital files are messy — full of placeholder values, obvious errors, and prices reported in the wrong units (we’ve seen a routine blood test listed at both $1 and $12,057). We run every price through a data-quality gate: statistical outlier checks, cross-checks against the Medicare rate and the hospital’s own list price, and machine-learning anomaly detection. Prices that fail are excluded from the BaseRate and from our comparisons. That’s why our numbers are often cleaner than the raw file — and it’s the whole point.