get_star_risk

Star Risk: odds a star falls, holds or rises next year. For one Medicare Advantage contract and origin rating year, for the Overall, Part C and Part D summaries, scored on the forecast year's own rules (the final rule's measure set and weights, named in the answer): Down / Same / Up for next year's star, the chance it falls below this year's star (chance to drop), the chance it lands below 4.0 (the quality bonus line; for a contract at 4.0 the same event and the same number as dropping, for one at 3.5 one minus the chance of rising) and the chance it rises (Upside), each with a status saying what it is worth. The Overall headline is Release 2r: a blend of Vlada's simulation (70%, Engine 4g) and a statistical model of the outcome with benefit and growth changes (30%), both calibrated on every year before the forecast year; the simulation's own odds ride beside it, labeled. Also: the expected star and its distribution (simulated), Star Distance (typical years of movement between the score before rounding and the line it earned), the raw score and the reward factor apart (and the chance the reward factor drops), three layers of distance (each measure to its own line, a cut point or for survey measures a modeled effective line; the score before rounding to its half-star lines; the raw mean and spread to the reward factor's thresholds), the Effective Star Cushion and Reach (descriptive), the cheapest sets of measure gains that reach the next half star (up paths), the per-contract Focus list (context) and the protect list (measures with the fewest score points above the line that keeps their star). With parent instead: the parent organization's book, the measures where contracts that improved moved up a star more often (expected members gaining a star, contracts helped; the top 3 are the held-out tested list), its protect list, and each of its contracts' Down / Same / Up.

ARGUMENTS: contract_id (e.g. H5216; H or R contracts, employer direct E contracts are not scored) or parent (a parent organization name such as Humana, or a contract id to use its parent); exactly one. part (Overall, C or D; default all three; contract only), rating_year (the origin rating year; default the newest scored), limit (rows, default 60).

RETURNS for a contract value.summary (contract, rating_year, forecast_year, engine version and runs, the headline model, the rule set, parts: one block per summary, each probability with a _status, down_same_up, and the simulation's own odds under simulation) and value.measures (one row per measure: star, weight this year and next, removed next year, weight share, distances down and up in score points and in typical years, the line they are measured to, lift per typical year); for a parent value.summary (parent, book, focus, protect) and value.contracts (each contract's stars and odds).

WHAT EACH NUMBER IS WORTH, as the answer says: on 2024-2027 (years read while the engine was built; RY2028 is the clean test), Release 2r's chance to drop ranks contracts well (AUC 0.771) and is sharper than history (Brier skill +0.079, 90% interval +0.058 to +0.101, against +0.032 for the previous model); its odds are close to calibrated (slope 1.08) but run about 3 points high on average (35.6% predicted, 32.8% actual); by origin star it predicted 14% Down for 3.0 contracts against 10% actual, 44% for 4.0 against 41%, 55% for 4.5 against 49% and 57% for 5.0 against 65%; a chance above 90% is shown as over 90% (only 5 such cases in history); a year-wide shock can move every contract together; below 4.0 is tested for contracts within half a star of the line (skill +0.100) and is context further from it; Upside is tested (from 3.5 it is the chance of reaching 4.0); Down / Same / Up is tested as one forecast; the simulation's own odds are shown only for comparison; Part C and Part D are simulated shares, uncalibrated and untested; the expected star is simulated and runs low; the parent focus list is an association tested on held-out years (top 3 only), never a promise of a star; the per-contract focus list did not pass its test and is context only; the protect list is a ranking rule, not a tested prediction, and never lists a measure CMS removes next year; slip paths are not served; the Effective Star Cushion and Reach are descriptive, never predictive; where the path engine does not reproduce CMS's published star its paths are flagged untrusted. Every number is a model forecast, not an observed rating.

SOURCE: CMS Part C and D Star Ratings Data Tables, Technical Notes, cut points and the forecast year's final rule, through Vlada's Star Risk engine (simulation on CMS's exact summary arithmetic, calibrated on every earlier year, blended with a statistical model fitted on every earlier year, judged on 2024-2027 and graded on RY2028 in October 2027).