The US Patent and Trademark Office granted Humana US 12,725,056 on September 1, 2026, for a system that forecasts time series by running a pool of machine learning models, scoring each against a chosen metric, and automatically selecting the winner. The application was filed on September 8, 2021, so the grant is a five-year-old idea arriving in a very different pricing environment.

In health insurance the time series that matters is trend. And trend is not a forecast an actuary produces. It is a selection an actuary signs.

Key Takeaways

  • The claim is model selection, not modelling. The system's contribution is the bake-off: run a pool of techniques, compute a metric for each, pick the best. The forecasting itself is off-the-shelf.
  • Selection on a backtest metric is the classic overfitting route. A model chosen because it scored best on history is chosen partly for having fitted that history's noise, and nothing in the claim describes a holdout discipline.
  • A winner that can change between periods destabilises the trend basis. If the selected model differs from one quarter to the next, consecutive trend picks are not drawn from the same method, and the period-over-period comparison a pricing actuary relies on stops being like for like.
  • "It won the bake-off" is not a rate filing answer. A regulator asking why this trend was selected needs a reason expressed in the mechanics of the book, not in a leaderboard position.
  • The environment makes this live rather than theoretical. Humana's Insurance segment benefit ratio hit 91.2% in Q2 2026 against 89.9% a year earlier, and the individual market missed trend by 9.9 points in 2025.

What the Claim Actually Covers

Read plainly, the patent describes a procedure rather than a model. The system determines a model metric for a specific use case, accesses a pool of machine learning models built on different techniques, forecasts with each, computes the metric for each, compares, and uses the winner for that application.

Every component is ordinary. Gradient boosting, state space models and the rest are decades of published work, and comparing candidates on an error metric is what any modelling team already does in a notebook.

What is being claimed is the automation of the choice, and the removal of the human from it.

That distinction matters more in insurance than in most industries that forecast. A demand planner who picks the wrong model discovers it in weeks and reforecasts. A health actuary who picks the wrong trend has embedded it in a rate that is filed, approved and in force for a year, and discovers it through the benefit ratio.

US 12,725,056Detail
AssigneeHumana Inc.
FiledSeptember 8, 2021
GrantedSeptember 1, 2026
Pendency4 years 11 months
Claimed contributionAutomated selection among a pool of forecasting models by metric comparison
Not claimedAny specific forecasting technique, or a holdout or stability discipline

Source: USPTO grant record, September 1, 2026.

Why Automated Selection Fights the Filing

Selecting a model because it scored best on historical data is how overfitting enters a process that believes it is being rigorous. The winning model wins partly on signal and partly on having accommodated the noise in that particular window. Nothing in the claim describes a holdout period, a stability requirement across windows, or a penalty for complexity.

The instability is the sharper problem. Suppose the selected technique changes between two consecutive experience periods because the metric ranked a different candidate first. Both forecasts may be individually defensible. The difference between them is not, because part of the movement is a change in method rather than a change in the book.

A pricing actuary reads consecutive trend estimates as a series and asks what moved. That question has no clean answer when the estimator itself moved.

Then there is the filing. A rate filing does not ask for a forecast, it asks for a justified assumption, and the justification has to survive a reviewer asking why this number rather than a lower one. An answer grounded in the book, that utilisation shifted in a named service category or a contract repriced on a known date, is reviewable. An answer that the model with the best backtest error produced it is not, because the reviewer cannot interrogate a ranking.

None of this makes the system unusable. It makes it a candidate generator whose output an actuary must then own, which is a narrower role than the claim implies.

The Five-Year Gap

The application was filed in September 2021 and granted in September 2026, a pendency of four years and eleven months. What was filed describes a pool of conventional techniques compared on an error metric, which was a reasonable 2021 description of automated machine learning and is not what a 2026 forecasting stack looks like.

That gap matters for reading any carrier patent as evidence of current capability. A grant establishes what was novel enough to claim when it was filed. It says nothing about whether the assignee still uses the approach, and the 2021 vintage of this one makes it weak evidence about how Humana forecasts trend today.

The environment has moved a long way in that window. Humana's Insurance segment benefit ratio reached 91.2% in the second quarter of 2026 against 89.9% a year earlier. Across the individual market, 2025 premiums grew 6.6% against medical expenses of 16.5%, a 9.9 point gap that no forecasting method selected in 2021 anticipated.

That is the useful way to read the grant. It is not a signal that carriers are automating trend selection, and it is not evidence that Humana has. It is a reminder that the automation of assumption-setting is patentable, which means it is being built, and that the accountability for a filed trend has not moved anywhere. It still sits with the person who signs it.

Further Reading

Sources