Corva Signal · MSSP Core
Understand why an ACO is projected to save.
Predict gross savings rate and surface the SHAP drivers that move the forecast up or down. Start with an MSSP ACO ID, then drill into the model signals behind the estimate in percentage points.
Every prediction comes with an explanation.
The model uses SHAP (SHapley Additive exPlanations) to decompose each predicted savings rate into individual feature contributions, expressed in percentage-point units. A SHAP value of +0.018 means that feature pushed the predicted rate up by 1.8 percentage points.
Analysis
Launch by ACO ID
Enter a CMS MSSP ACO identifier and receive a predicted gross savings rate with key driver attribution in seconds.
Estimates are produced by a model trained on historical CMS MSSP data (2017–2023). They are not clinical or billing guidance and are not a guarantee of future performance.
Model Validation & Methodology
The MSSP Core uses a machine learning model called gradient boosting, which builds predictions by systematically combining hundreds of decision rules learned from historical data. The model learns the relationship between network characteristics and gross savings rates from historical MSSP results, then applies those learned patterns to predict the performance of any network configuration.
To ensure the predictions hold up in the real world, not just on historical data, the model is validated exclusively on a future year it never trained on (PY2023), which is the actuarial equivalent of testing a forecast against outcomes that hadn't yet happened when the model was built.
The model predicts gross savings rates within an average of 1.94 percentage points, explaining approximately 59% of the variance in ACO performance — a meaningful signal in a domain where outcomes are driven by dozens of interacting clinical, behavioral, and regional factors that no single metric can fully capture.
Mean absolute error
1.94pp
On PY2023 held-out test set — never seen during training
R² explained variance
0.593
Out-of-time validation, PY2023 holdout
Training universe
3,350
ACO-year observations, 1,241 unique ACOs
Years of history
7yrs
PY2017 – PY2023, spanning pre- and post-COVID benchmarks

