Employer Health Insurance Renewal Prediction With the Benefitra Renewal Index

1 Why renewal prediction is hard

What data can actually answer the question boils down to two main differences.

Renewals look simple. It's just one number, the percentage by which next year's health insurance premium beats this year's. But in reality it comes at the end of a long chain, fed by unit price, utilization, service mix, plan design, demographics, area cost, carrier margin and regulation. Each of those links shows up in different data, and the data you end up using depends on how the group is rated.

Where the small group line sits

Federal rules define the small group market as employers with 1 to 50 employees. States are allowed to stretch that to 100, and California, New York and Vermont actually do that. Premiums inside this market can vary based only on age, area, tobacco use and family size. They never vary according to health status. New York and Vermont take it even further and do not allow age rating at all. There's also the Applicable Large Employer threshold of 50 full time equivalent employees. This threshold controls the employer mandate but does not control the rating method. So a group could be an ALE for the mandate and still be community rated in a 1 to 100 state. When it comes to renewal prediction, there is one important question to ask. Is this group's price filed and readable, or is it experience rated and negotiated?

2 The landscape of renewal signals

By segment, here is what really tells you what next year's renewal will be.

Lots of organizations publish a medical trend, a cost index or a premium survey. But they are not all equally useful when you are trying to predict a renewal. Below, we cite vendor surveys as single context figures with dates. We do not reproduce their tables.

The three figures don't measure the same thing. PwC's is a forward medical cost trend for 2027. KFF's is what employers paid in 2025. Altarum's cover economy wide health prices and usage as of mid 2026. Not one of them is a renewal forecast for a specific group.

Small group signals, ranked

Large group signals

Large groups have no public filings. What you can use is mainly the group's own claims experience, plus national medical trend surveys for context and public area cost proxies such as BLS metro price series and CMS geographic cost indices. We don't publish a large group forecasting model in this paper. A sound one would take licensed inputs, or else a rebuild on free sources like MEPS-IC and CMS National Health Expenditure data.^{5, 18}

3 Five year outlook

Over many years, a forecast is a decomposition and not an index.

Health economists split nominal medical trend into inflation, real income growth and "excess" growth, and over the long run they force it toward the growth of the economy.

Why does this matter? Healthcare spending cannot grow faster than the economy forever. So a credible five year forecast takes a hot near term rate and bends it downward, rather than extending it.

Public long run instruments

Remember that the CMS private health insurance figures are aggregates, and shifts in enrollment keep them down. Growth per enrollee (the renewal concept) is higher, so think of the aggregate as a floor.

A five year glide

Our five year recipe has four steps. First, we anchor the long run at the Getzen figure of about 5.0%. Next, we set years 1 and 2 at about 7 to 8% per member and glide between them. Last, we flex by state by plus or minus 1 to 1.5 points. The near term 7 to 8% is a Benefitra planning estimate, and we have not tied it to any one public source. On average, the five points in Figure 1 come to 6.5% a year.

Figure 1. Five year national glide path (Benefitra estimate). Year 1 is the first projection year. The points are 7.5, 7.1, 6.5, 5.9 and 5.4%, and they average 6.5%. The band is a constant ±1 point each year, and it is not compounded. Year 1 is below PwC's 9% forward medical cost trend for 2027, but the two measure different things and we have not reconciled them. Figures are per enrollee, before health plan buydowns.

4 What a rate filing tells you

In small group, the filed rate answers the question for the year that it covers. The filing does not forecast any later year.

Each year, individual and small group carriers file their rate changes through the federal rate review process.^{9, 10} Review kicks in at an increase of 15% or more, or one above a state specific threshold.¹¹ Large group generally doesn't go through federal rate review.¹²

Requested rates did not help

Does a carrier's requested rate help predict that carrier's rate change the next year? What about the gap between what carriers ask for and what actually gets approved? We checked both of these using the CMS rate review files for plan years 2014 through 2026 (Section 5B).²¹ It turns out that neither one helps. The requested rate does not add anything to next year prediction (ΔRMSE +0.013 percentage points, n = 4,473). As for the gap between the requested and the approved rate, it has close to zero correlation with the year that follows. We haven't looked at whether that gap matters in prior approval states. The regulator modified field appears only in the files from 2014 to 2019 and covers about 3% of issuer years.

If you have the approved rate, it's the better anchor for a small group block. It isn't in the CMS public files after 2019, though, so you have to get it from RateReview.HealthCare.gov or the state insurance department.¹⁰ One context point on the individual market (not group). Peterson-KFF found a median proposed 2026 increase of 18% across 312 Marketplace insurers.¹³

5 Backtests of which lookback predicts most accurately

If all you have is past data, which rule holds up when you test it on new data?

Every model needs a base rule for when there is no filing. If all you have is recent history, what gives the most accurate forecast of next year? We tested this first on a national series and then on the CMS rate filing panel. The evidence from the CMS panel is stronger.

The national series and its data problem

For our national test we used the KFF Employer Health Benefits Survey data on the all firms family premium annual increase, from 2000 to 2025 (a total of 26 observations). This series was rebuilt using published KFF cumulative figures, and the interior years are estimates. We have left the swap to MEPS-IC data for a later update. Until that happens, the national results in this edition rely on the reconstructed series.¹⁸ In an earlier draft, an AI model filled in those interior years with values it had recalled. We do not present those as data. The only values we have confirmed against KFF's own pages are recent ones. They show +7% for 2024 and +6% for 2025.^{2, 3}

We ran a spot check. Over roughly five year windows (the first one, 1999 to 2005, spans six years), we compared the growth implied by the reconstructed series with KFF's published average family premiums (KFF 2025 survey, Figure 1.12, and its stated five and ten year growth).²

In four out of the five windows, the reconstructed series lands within about 1.5 points of the totals KFF published. In 2005 to 2010 it runs 3.7 points high. Is that close enough to give a rough test of how the rules rank? Yes. Is it close enough that we could publish the series as data? No, and so we don't publish it.

Method

For each test year from 2008 to 2025 (that's 18 years), each rule gets fit only on data from the years before and then it has to predict that year. Because of that, the number of training points grows from 8 to 25 as we work our way through the years. Each rule got scored on root mean square error (RMSE, where big misses count more) and also on mean absolute error (MAE). In total there were 14 rules that we looked at. These included naive last year, moving averages, exponential smoothers (EWMA), fitted autoregressions and the long run mean. Keep in mind that this is a retrospective backtest. We have not yet frozen a forecast and scored it prospectively.

Throughout this paper the EWMA uses equal weight (α = 0.5). The estimate for next year = 0.5 × the latest year + 0.5 × the previous estimate. That means past years get successive weights of 0.5, 0.25, 0.125 and so on.

One more, for reference. On the same test, the mean of the last 6 years gets RMSE 2.54 and MAE 2.12. It was not part of the original 14 rule set.

Figure 2. KFF all firms family premium annual increase, 2000 to 2025 (national, all firm sizes). The top panel shows the reconstructed series with the equal weight EWMA forecast, and the shaded years mark the out of sample test. Below it is the RMSE for all 14 rules. Keep in mind this isn't a small group series, and the interior years are estimated.

What the national test does and does not show

Lowest RMSE only. EWMA 0.5 did have the lowest RMSE. MAE is another story. Naive last year tied it (1.761 vs 1.762), and three other rules beat it, the 0.5 / 0.5 blend (1.709), EWMA α = 0.7 (1.730) and the 0.3 / 0.7 blend (1.755).

Not statistically significant against naive. If you run a Diebold Mariano test on squared errors, you get t ≈ -1.15 on 18 points. If you take out the 2011 spike and its 2012 rebound, the two rules score the same (RMSE 1.945 vs 1.946).

The supportable claim is modest. On this series, reasonable smoothing does at least as well as naive last year and a lot better than the long run mean. Part of why the long run mean does so badly is a regime effect, because the double digit years of 2000 to 2005 stay in its window.

Proxy. The all firms family premium leans toward large and self funded employers. It is not a small group book.

Lookback window for the five year horizon

Five year paths raise a different question. If you want to predict the average over the next five years, how many years of history should you average? The draft pushed the national series back to 1988 (KFF/KPMG HIAA, with values before 2000 estimated and 1989 to 1998 at low confidence) and it tested windows of 1 to 20 years. The same data caveat applies here, and even more strongly.²²

Figure 3. RMSE by lookback window length, for next year versus five years ahead (low confidence). Grey shows the next year target. Green shows five years ahead on the 2008 to 2025 origins, and navy dashed shows five years ahead on a longer set that also includes the volatile 1989 to 2004 origins. The amber line marks the six year window. The data is the national series from 1988 to 2025, partly reconstructed. Values before 2000 are estimated, and there are about three independent five year windows. On the chart, the "optimal" labels mark only the minimum of each curve.

Three things stand out in the numbers.

Recent calm period (2008 to 2025 origins). The three year window had the lowest error (RMSE about 1.24). The six year window scored 2.00, and the 20 year window scored 2.66.

Longer set including the 1990s and early 2000s. Here the error rises from about 4.0 at one year to about 4.7 at four years. After that it falls to a low of 2.41 at 15 years and climbs again to 2.66 at 20 years. The six year window scored 4.46. That makes it worse than the two year window (4.33), slightly better than the three year (4.61) and much worse than 15 or 20 years.

No single optimal window. Which window has the lowest error changes sharply from one set to the other. And the test rests on about three independent five year windows plus estimated values before 2000. It doesn't identify an optimum. It doesn't validate the six year choice either.

5A. Carrier versus state when one carrier's filing is missing

Here is the practical question. If one carrier's filing is missing, should you fill it from that carrier's own history, or from the smoothed path of the state? We tested both approaches on the CMS rate review panel.²⁰

Data. CMS Single Risk Pool Rate Review Public Use Files, Worksheet II, plan years 2014 to 2026, small group only.⁹ Where there is a regulator modified rate, we use that one. If there isn't one, we fall back to the submitted rate, then the product rate, then the cumulative change. We drop changes of more than 50 points. Issuers need at least three years to stay in, and everything is weighted by enrollment. In total that covers 50 states and DC. There are 630 issuers, 5,987 eligible issuer years, 3,964 issuer test observations and 454 state test observations. The test years run from 2018 to 2026, and each rule is fit using only the years that came before it. For the most part the outcome is the submitted rate. Of the 6,268 raw issuer years, 3,066 used the submitted field and 2,979 used the product field. Another 213 used the regulator modified field, and 10 used the cumulative field.

State grain

The top rules are close together (4.39 to 4.56). Naive last year stands out as clearly the worst.

Issuer grain

The rules that were based on the state came out ahead of every one of the carrier's own rules. Compared with the carrier's own EWMA and its three year mean, the margin was 0.75 to 0.92 RMSE points. The two leaders differ by less than 0.01 and are effectively tied. There is one result, though, that goes against the national test. At the issuer grain, the carrier's own long run mean (7.191) and six year mean (7.175) both did better than the carrier's own short EWMA (7.236). Carrier histories are noisy enough that the longer averages help. That's why we don't say "never use a long run mean". What we say is "prefer the state path".

Figure 4. Issuer grain RMSE (blue) and MAE (orange), n = 3,964. The lowest scores go to the state transfer and light blend rules, and the more carrier weight you add, the higher the error gets. Labels are rule codes. "ewma_0.5" is the carrier's own EWMA, "state_ewma_0.5" applies the state path, and "shrink_0.25" is the 25/75 blend. CMS rate review data, small group, 2018 to 2026 test years.

Figure 5. RMSE by rule at issuer grain and at state aggregate. State error (about 4.4 to 5.0 points) is well under issuer error (about 7.0 to 8.5 points). Same panel.

Why the state wins

From one year to the next, carrier series barely persist (the AR(1) coefficient is about 0.15, compared with about 0.27 for state series). That is why it helps to borrow strength from the state. Even within the same state, insurance carriers differ a lot. Across state years that had two or more carriers (n = 652), the median spread between the 25th and 75th percentile carrier was 4.93 points. In 8 of 9 test years, the state rule with the lowest error was not the same as the issuer rule with the lowest error. So no, the state average can't take the place of a carrier's filing. It is just the better default to use when that filing is missing.

Figure 6. Median spread of carrier rate changes within each state, by filing year. It shows the median interquartile range and the standard deviation across state years that have two or more carriers. Same panel.

5B. Requested rates, a dead end we tested

These are the same CMS rate review files, but the sample is wider than in Table 6 (n = 4,473 against n = 3,964). So the baseline here is different from Table 6.²¹ We added the carrier's requested rate for year t to a regression on the state EWMA. That improved t+1 RMSE by 0.013 points (6.686 to 6.673, n = 4,473), which is well below a usefulness bar of 0.05 points. At t+2 it made things slightly worse (-0.004, n = 3,876). If you carry the requested rate forward on its own, it scores an RMSE of 8.722 at t+1, which is far worse than carrying the state EWMA forward (6.756). The correlation with the next year was 0.20 for the state EWMA and 0.14 for the requested rate. For the gap between requested and approved, it was about 0.00. The regulator modified field only exists in the files from 2014 to 2019. So any claim about prior approval states is still untested.

So the bottom line is that the requested rate adds nothing to next year prediction. To fill next year gaps, use the state EWMA or the 25/75 blend.

5C. Modeled Taft-Hartley (Benefitra modeling assumption)

The Taft-Hartley plan is a national multi employer medical plan on the BCBS PPO network. Its trust sets its pricing. That makes the state small group filed averages in Section 5D the wrong input for it, so Benefitra models it on its own path.²⁴ To see how these plans work for employers, read Taft-Hartley plans for small employers.

That 1.5 point SD is the forecast uncertainty used in the simulation, and it is wider than the spread of the path itself (about 0.5).

In the stress case we add about 2 points a year to the path from 2028 onward. The averages come out to around 10.0% over three years and 9.5% over five years.

This is for comparison only, and it isn't like for like because the plan is national. 7.5% is higher than 27 of the 51 state six year averages in Section 5D (median 7.36%). And 8.0% is higher than 15 of the 51 state three year averages (median 9.42%).

5D. State small group averages, 2021 to 2026

These are small group filed rate changes by state, weighted by enrollment, taken from the CMS rate review files.^{9, 23} The state fully insured defaults in the simulator should use this vintage of six year averages.

The median six year average comes to 7.36%, and the median three year average to 9.42%. Most states saw the last three years run hotter than their six year average.

5E. The Benefitra Renewal Index

The method that the backtests pick is what we call the Benefitra Renewal Index. It is an open rule for estimating a small group health insurance renewal when you don't have a live filing at hand. It is not an industry standard.

Open formula

Start with y, a state's enrollment weighted small group filed change. The formula has three parts.

How it compares

We don't score PwC, Segal or other survey trends here. Doing that would require survey forecasts that were locked at the time they were made, and we don't have those. We looked in a 2026-09-18 scan. We did not find any public consulting, filing template or trade group product that publishes this exact open recipe. The closest relatives are the multiyear experience weighting and exponential smoothing methods that appear in actuarial education papers.

6 Specialty drugs

A new specialty drug causes a shift in cost trend. Its biology tells you what shape that shift will take.

A drug that cures works like a one time pulse. Every patient it treats leaves the eligible pool, and on top of that competition lowers the net price, so spending goes up at first and then comes back down again. A chronic therapy works more like a wave, because patients keep on taking it. Spending keeps climbing until it reaches a plateau, and it only bends with price cuts, patent expiry or tighter eligibility. In a renewal model these two shapes need opposite treatment. If you treat a pulse as a permanent level, you will over price every renewal that comes after it.

Figure 7. Pulse versus wave (a stylized schematic, not data), comparing Hepatitis C (a cure) with GLP-1 therapy (chronic). The "-80% by 2018" annotation is based on Gilead Hepatitis C revenue figures that have not yet been checked again against Gilead's annual reports. Treat it as an illustration.

Hepatitis C as a pulse

Gilead's Sovaldi was approved in December 2013, and Harvoni followed in October 2014. Spending spiked. Then it fell, as the pool of treatable patients shrank and competing prescription drugs cut net prices. Gilead's own filings describe revenue falling as the number of suitable patients declined. (The year by year figures still have to be pulled again from Gilead 10-K filings.) The lesson? A cure adds a one time loading and then turns into a tailwind.

GLP-1 therapy as a wave

GLP-1 drugs are used long term. The eligible population is large and there's no pool drawdown. IQVIA reports that GIP/GLP-1 agonists contributed $14 billion of US net spending growth in 2025.¹⁶ KFF's 2025 employer survey tracks rising GLP-1 coverage among large employers, along with their cost concerns.² Oral versions, patent expiry and price competition may bend the slope. But they don't shrink the eligible pool.

Early indicators (qualitative)

An earlier draft had a chart that rated these signals by "predictive strength". There was no data behind those ratings, so we removed the chart. Does the prescription drug pipeline make renewal forecasts better? That is a hypothesis we have not tested.

A proposed drug loading

The proposed model adds each major drug to the baseline cost trend as its own named, per member loading. That way the pulse can be taken out as it decays. Here's how the steps go. Classify the drug as a cure or as chronic. Then estimate the eligible population. For a cure that is a shrinking stock, and for a chronic therapy it is a growing one. Apply an uptake curve, and then a net price path that steps down when competitors launch and when the patent expires. Finally, divide by expected claims.

We haven't fitted or tested this loading.

7 A proposed composite

Standard trend plus one overlay. It's a design. It isn't a result.

There's no accepted academic index that signals commercial renewals ahead of time. The closest public composites are the CMS NHE projections, Altarum's monthly indicators and the Peterson-KFF tracker.^{5, 8, 13} We would rather not invent an index. What we propose instead is a standard base with one overlay added on top, picked from three candidates.

A. Pharmacy pipeline overlay (1 and 5 years). The Section 6 drug loading, on top of a base from CMS NHE or from published cost trend surveys.

B. Stop loss and high cost claimant loading (self funded, 1 year). Built from the group's own stop loss quotes.

C. Labor and macro anchor (5 years). It uses BLS health care wage data and hospital prices, and it pulls the long run toward GDP growth.

The overlay weights ought to come from the data rather than being set by hand. And so far, none of these overlays has been fitted or backtested.

8 Recommended architecture

9 Does the renewal assumption change the funding recommendation?

Method to be repeated

In the funding simulator, the default fully insured renewal is a filed average for each state. It only reaches community rated groups (up to 50 lives in most states, and 100 in CA, NY and VT). Experience rated groups renew on their own cost trend and do not read it. The test compares two arms across 35 segments ({NY, CA, MA, FL, TX, AL, WY} × {30, 70, 150, 250, 500} lives). One arm uses the per state default. The other uses a flat 6.75% for every state. For each segment, the test records the five year probability that each funding path beats fully insured, and also whether the recommended path changes.

Four points have to be handled in the rerun.

State vintage. Use the 2021 to 2026 averages in Section 5D. They differ from the values used before. NY is 9.36 (was 9.4), MA 7.67 (7.4), TX 7.66 (7.3), FL 6.82 (5.9), AL 5.48 (5.4), WY 5.28 (5.3) and CA 5.44 (5.0). FL now sits above the 6.75% flat index. That reverses its direction compared with the earlier run.

New York. In New York, self funded, level funded or self funded captive arrangements require 101 or more full time equivalent employees. Groups of 100 or fewer are small group in New York and can't use them. Their options are fully insured, PEO, Taft-Hartley and MEWA. So those paths must be excluded for NY 30 and NY 70.

Construction. By design, segments that are experience rated cannot move. The report is going to say how many segments could move at all, and near the 70% threshold it will give exact probabilities.

Conflict of interest. Benefitra sells a number of the alternative paths, and the Taft-Hartley plan is one of them. Any recommendation of a path will come with that disclosure.

Suppose the recommendation stays the same when one input is swapped. That would only show that the recommendation is not sensitive to that input. It would not show that the recommendation is right.

10 Methods and data sources

Backtest design. The backtests are rolling and retrospective, and they are out of sample. Each rule only gets fit on data from before the year that it predicts. The metrics are RMSE and MAE, both in percentage points. Section 5A gives the CMS panel field rules, drop rules and weights. The EWMA starts from the first observed value. Since we picked the rule set and α = 0.5 after we had already seen the same data, these results are not a prospective test. We don't report confidence intervals yet, except for the significance test in Section 5.

11 Limitations

• The national series is reconstructed. Its interior years are estimated, and the values before 2000 are even more so. Until we rebuild the series on MEPS-IC, Table 4 and Figures 2 and 3 should be treated as indicative only.
• The national results aren't significant. On 18 test points, EWMA doesn't beat naive last year by a significant margin.
• The five year window has not been tested as an optimum, and the direct test on CMS and MEPS-IC is not complete yet.
• For the CMS outcome, what we mostly have is the submitted rate, not the approved one. After 2019 the approved rates are missing from the public files.
• It's retrospective only. No forecast has been frozen in advance and scored against finals.
• Large group isn't modeled. There's no public filing panel.
• Estimates. Both the five year glide and the near term 7 to 8% are Benefitra estimates, and the Taft-Hartley path after 2027 is a Benefitra modeling assumption.
• Not reverified in this edition. The BLS and CBO pages block automated access, so the figures that rest on them were not rechecked. The yearly Hepatitis C figures also still need to be pulled again from Gilead's filings.
• The state detail covers small group only. Do not apply it to larger groups.

12 Data, citation and corrections

Download the data

If you want the Figure 3 lookback series, the national series and the backtest code, you can ask us for them.

In the web edition, your browser builds each CSV from the matching table and adds a source line.

License. © 2026 Benefitra. Text and Benefitra derived tables are licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). You may share and adapt with credit to Benefitra and a link to this page. Third party data (CMS, AHRQ, KFF, vendor figures) remains under its owners' terms. The Funding Simulator software is proprietary and not licensed for reuse.

How to cite

Newland, S. (2026). Employer Health Insurance Renewal Prediction With the Benefitra Renewal Index (White Paper No. 01, v1.0.2 preview). Benefitra. https://benefitra.com/renewal-index/.

Version policy

The Index has a version number. In v1.0, the near term uses EWMA α = 0.5 on state small group filed changes. Five year paths use the six year state average, and carrier gaps use the 25/75 blend. If we ever change the method, the change will come with a written reason and backtest evidence. There will also be a notice on this page and an entry in the changelog.

Changelog and errata

Spotted an error? Email info@benefitra.com or contact Benefitra. We log every correction here, with the date, what changed and which numbers moved.

13 References

Questions employers ask about renewal prediction