How a Medicare Advantage risk score is built

CHAPTER 01

What the number is

Every person in Medicare Advantage carries a risk score. The scale is anchored to one fact: the average person in Original Medicare scores 1.0. A member expected to cost half the average scores near 0.5, one expected to cost double scores near 2.0, and the plan's payment moves with it dollar for dollar. The score is not a judgment about the member. It is a prediction of cost, computed from two ingredients: who the member is, and what conditions their diagnoses establish.

Start with who they are. Age and sex set a base rate. Then the model asks which of its seven population segments the member belongs to, because the same condition costs different amounts in different populations: living in the community or in an institution, entitled by age or by disability, and, in the community, whether the member also has Medicaid, in part or in full. The dual-eligible seam that Payer Pulse Report 03 maps at market scale starts here, in the coefficients.

CHAPTER 02

A diagnosis is not a payment

The model does not pay for diagnoses. It pays for Hierarchical Condition Categories: groups of diagnoses that predict cost together. The 2026 model maps 8,019 ICD-10 codes into 115 payment categories, and the path from a code on a claim to a coefficient in the score has three places where the money can change or vanish on the way through.

First, most codes map to nothing. ICD-10 has tens of thousands of billable codes and the model pays on 8,019 of them. Hypertension, high cholesterol, and coronary artery disease without angina, three of the most common diagnoses in Medicare, are simply not in the 2026 mapping. Second, a hierarchy: when a member has both a severe and a milder form of the same condition, the severe one is paid and the milder one is suppressed to zero, so coding both adds nothing. Third, eleven interactions: certain pairs, like diabetes together with heart failure, predict more cost than the two alone, and the model adds a coefficient when both are present.

CHAPTER 03

Raw is not paid

The number the model produces is not the number CMS pays on. Two adjustments sit between them, both set each year in the Rate Announcement, and public discussion of risk adjustment almost always quotes the raw score as if they did not exist.

The first is normalization. Diagnosis coding rises over time even in Original Medicare, so a model calibrated on old data would slowly drift the "average" member above 1.0. CMS divides every score by a factor that puts the average fee-for-service beneficiary back at exactly 1.0 for the payment year. For 2026 that factor is .

The second is the coding pattern adjustment. MA plans document diagnoses more completely than fee-for-service Medicare does, so statute requires CMS to cut every MA risk score by at least 5.9 percent to offset it. For 2026 CMS applied exactly that statutory minimum, as it has every year since the floor was set.

CHAPTER 04

The multiplier meets the county

A risk score has no dollar value of its own. It multiplies whatever the county machinery produces, and that machinery is a different page: The Rate Book builds the county benchmark from the fee-for-service base, the quartile percentage, and the star bonus, and The Star Book explains the rating that switches the bonus on. Hold all of that fixed and the risk score is the last multiplication: the same member is worth different dollars in different counties, and different members are worth different dollars in the same county.

The counties below are real rows from the 2026 ratebook, the same audited data The Rate Book runs on. The spread is the point: the no-bonus benchmark runs from $649.93 in Adjuntas, Puerto Rico to $2,536.98 in North Slope, Alaska. Same model, same member, almost four times the dollars.