METRA Workforce Claims Forecast — Methodology v2.1

The Experience Calibration addendum. It specifies how a group's own paid medical and Rx experience is credibility-blended against a manual rate, and how the forward metabolic signal from the Workforce Exposure Forecast (Methodology v1.0) is then applied as a dollar offset to return a projected forward claims figure. v2.1 hardens the band: the credibility constant is re-derived as an independent limited-fluctuation full-credibility anchor (§6), and the single inherited band is replaced by two labeled bands from one variance model — an expected-claims confidence band and a realized-paid predictive interval (§9). Published in the open, in v2.1, for review and attestation by credentialed actuaries.

§1. Scope and Intent

What this addendum specifies. The procedure by which Metra calibrates a single employer cohort's forward claims expectation using that cohort's own paid medical and pharmacy experience. The procedure has three stages: (a) mature and pool the connected paid experience into a credibility base; (b) blend the resulting observed per-employee-per-year (PEPY) rate against a manual or expected rate using Bühlmann credibility; and (c) project the blended rate forward one year under a medical-trend and attrition assumption, then subtract the forward metabolic dollar signal produced by the v1.0 Workforce Exposure Forecast. The output is a projected forward claims figure with a confidence band.

What this addendum does not specify. It does not specify a rate filing, a stop-loss credibility table, or a replacement for the carrier's own renewal math. It is the buyer-side counterpart to that math: the same renewal arithmetic a carrier already runs on the employer's population, run by the employer, with the forward metabolic signal folded in as a disclosed, auditable offset. The output is a forecast the CFO can plan against and the actuary across the table can check; it is not raw material handed to the reader to build their own model.

Relationship to v1.0. v1.0 is consumed here as a finished input. The 12-month metabolic point estimate and its confidence interval enter this procedure as fixed dollar quantities. v2.0 does not re-open, re-weight, or re-credibilize the v1.0 result; it nets it against the gross claims projection.

§2. Notation

Symbols introduced in this addendum. Where a symbol also appears in v1.0, the v1.0 definition governs unless noted.

§3. Connected Experience and Data Quality

The procedure consumes aggregate paid experience only: total paid medical, total paid Rx, and the enrolled member-months that generated them, for a defined incurred period with a stated paid-through date. No member-level claim records, diagnosis codes, or any other protected health information enter the procedure (see §10). The connected experience carries a source label (carrier report, TPA file, ASO summary) so that the provenance of the calibration is recorded with the output.

Data sufficiency is the actuary's first checkpoint. The procedure does not impute missing months, does not annualize a partial period by assumption, and does not silently substitute a benchmark when experience is thin — it surfaces what was supplied and what was defaulted. Where the group has not supplied a manual rate, the expected rate falls back to a sourced, dated national benchmark — the average total health benefit cost per employee from Mercer National Survey of Employer-Sponsored Health Plans, 2024 ($16,501) — which is carried verbatim into the output (with its source and year) and explicitly flagged as a benchmark to be replaced with the carrier manual rate. This benchmark is refreshed when Mercer publishes a new survey. A blend resting on the benchmark is weaker than a blend resting on the group's own manual rate, and the output says so.

§4. Completion and IBNR Maturation

Paid claims for a recent period are immature: claims incurred near the end of the period have not yet been paid. The procedure matures paid to an incurred basis in one of two mutually exclusive ways, in priority order:

When neither is supplied, C = P and the experience is treated as fully mature — appropriate only for a run-out-complete period. The completion factor is the actuary's lever; it is supplied per group, not assumed by the platform, and it materially moves the observed rate. The procedure rejects a non-positive completion factor.

§5. Large-Claim Pooling

A single catastrophic claimant can dominate a mid-size group's experience and destroy the credibility of the observed rate. Where a pooling point is supplied, the dollars above it are excluded from the credibility base:

The pooling point itself is informational and the excluded dollars X are reported alongside the result, so a reviewer can see how much of the raw experience was set aside and reconstruct the unpooled figure. Pooling is applied to the credibility base only; it does not alter the manual/expected rate, which is assumed to already sit on whatever basis the carrier quoted. Pooling is optional — when no excess is supplied, B = C and the result is flagged as an unpooled basis.

§6. Credibility Blend of Observed and Expected

The observed PEPY is the pooled, completed base per average enrolled life:

It is blended against the expected/manual PEPY using the Bühlmann credibility weight:

The interpretation is the standard renewal one: a small group (low N) receives a low Z and its forward rate leans on the manual; a large group (high N) receives a high Z and its forward rate leans on its own experience. At N = k = 400 the weight is exactly 0.5.

Where k = 400 comes from (v2.1). The constant is set as a limited-fluctuation full-credibility anchor, derived independently of v1.0 rather than borrowed from it. Fix the standard for full credibility at Z* = 0.90, reached at N* = 3,600 life-years of exposure — a conventional group-health full-credibility threshold. The Bühlmann constant that places the partial-credibility curve through that anchor is:

So 400 is not a convenience constant; it is the credibility constant implied by a stated full-credibility standard. The same value happens to govern v1.0's metabolic weight, but the two are now justified separately, on their own exposure bases — a reviewer may reset either anchor without touching the other.

Implied EPV / VHM — a reasonableness check, not an empirical fit. Bühlmann's constant is k = EPV / VHM, the ratio of the expected process variance per life to the variance of the hypothetical means. Adopting the between-employer dispersion τ (§9) as the VHM scale, a = (τ·E)², the constant implies a per-life process variance EPV = k·a — i.e. a per-life process SD of s = √k·τ·E. At k = 400 and τ = 0.15 this is an implied per-life pooled-claims CV of √400·0.15 = 3.0:

A per-life CV near 3× is entirely ordinary for a single year of one person's pooled medical + Rx spend — most lives spend little, a few spend a great deal. The check is presented to show the constant is not in conflict with the dispersion the band assumes; it is not an empirical estimate of EPV and VHM from member-level data, which the aggregate-only procedure (§3, §11) never touches. The reviewing actuary substitutes the block's own dispersion through τ.

§7. Forward Projection and the Metabolic Offset

The blended rate is carried forward one year on the surviving lives, trended, then reduced by the forward metabolic dollar signal:

Here M is the v1.0 Workforce Exposure Forecast 12-month point estimate — the modelled forward healthcare exposure that the cohort's measured metabolic trajectory is expected to avoid over the year. It enters as finished dollars. The trend t defaults to a conservative single-digit anchor and is intended to be overridden with the carrier's own renewal trend; the attrition α defaults to the v1.0 value. The implied trend impact of the metabolic offset, M / G, is reported so the reader can read the offset as a fraction of gross.

§8. Non-Double-Counting of Credibility

This system contains two distinct credibility applications, applied to two different quantities, and they are never stacked. This is the single most important integrity property of the addendum and the first thing a reviewing actuary should verify.

The metabolic offset M arrives already credibility-weighted by v1.0's Z_metabolic. v2.0 consumes it as a fixed dollar figure and applies no further credibility weight to it. Z_claims touches only the observed-versus-expected PEPY blend and never touches M. The two weights operate on disjoint quantities; there is no path by which a single dollar of exposure is credibility-discounted twice.

§9. Confidence Band and Predictive Interval

v2.0 inherited its band entirely from the v1.0 offset CI and treated the gross projection G as a point. v2.1 replaces that with two labeled bands built from a single Bühlmann–Straub variance model, so the figure carries both the uncertainty in the estimated mean and the volatility of realized claims. The model has one disclosed dispersion input, τ, the between-employer CV of the true mean PEPY (default 0.15, overridable), from which the VHM is a = (τ·E)² and the per-life process SD is s = √(k·a) = √k·τ·E.

(1) Expected-claims confidence band. This is the headline band — the estimation uncertainty of the mean forward claims, combining two independent sources in quadrature: the credibility estimation error of the blended PEPY, and the estimation error of the metabolic offset (from its own 95% CI).

(2) Predictive interval for realized paid claims. A renewal is not only an estimate of the mean — next year's realized pooled claims fluctuate around that mean. The predictive interval adds the forward-year process variance of the retained block to the estimation band. It is labeled separately and reported pre-risk-charge: it is the statistical spread of paid claims, not a priced premium, and it excludes the stop-loss layer above the pooling point.

Both bands are centred on the projected figure Ĝ and are symmetric. The expected-claims band answers “how well do we know the expected cost”; the predictive interval answers “how wide could the realized paid claims land.” The predictive interval is always the wider of the two. A reviewer who wishes to stress trend, completion, or dispersion can do so directly through the disclosed inputs t, c, and τ. Both bands are reported on-screen and in the signed Workforce Claims Forecast PDF.

§10. Worked Example

A 322-life group with a full plan year of connected experience, a 0.92 completion factor, a $150,000 pooling point with $380,000 above it, a carrier manual rate of $16,500 PEPY, an 8% forward trend, 15% attrition, the default between-employer dispersion τ = 0.15, and a v1.0 metabolic offset of $240,000 (CI $160,000–$330,000).

§11. PHI, Aggregation, and Governance

The calibration is aggregate-only by construction. The connected-experience record stores period-level dollar totals and member-months — never a member-level claim line, diagnosis, procedure, or identifier. The forward metabolic signal it consumes is itself a cohort-level figure subject to the v1.0 small-cell governance floor. The calibration therefore never touches, joins, or re-identifies protected health information, and there is no path in the procedure by which a single member's claim or biometric history could be reconstructed from the output.

Access to the calibration and to the signed output is gated to the employer's forecast-authorized administrators (the primary administrator and HR administrators). Every generated calibration is recorded in an audit trail carrying the inputs' provenance, the resulting credibility weight, the methodology version, and a reproducibility digest of the calibration inputs, so any signed figure can be regenerated and checked.

§12. Conformance with Actuarial Standards of Practice

§13. Limitations and Refusal Conditions

1. The blend is only as good as the manual rate.

   When no group manual rate is supplied, the expected term falls back to a national benchmark and the output is flagged. A benchmark-anchored blend is materially weaker than a manual-anchored one and should not be carried to a renewal table without the carrier's own manual rate.

2. Completion and trend are supplied, not derived.

   The procedure does not estimate a completion factor or a trend from the data; it applies what the actuary supplies. A wrong completion factor moves the observed rate proportionally and a wrong trend moves the entire gross projection.

3. Experience volatility is now in the band — within stated scope (resolved in v2.1).

   v2.0's single band reflected only metabolic uncertainty. v2.1 adds an expected-claims confidence band and a realized-paid predictive interval from the §9 variance model, so experience volatility is now quantified through τ. What the predictive interval still excludes is stated, not hidden: it is pre-risk-charge, it does not re-add the stop-loss layer above the pooling point, and it does not price a carrier margin. The dispersion τ is a disclosed assumption, not an empirical member-level estimate.

4. The metabolic offset is inherited from v1.0 and may double-count across markers.

   v1.0 monetizes weight, blood pressure, blood glucose, and waist on additive per-marker coefficients. When one physiological improvement moves several markers at once, the additive offset can overstate avoided exposure. Because the offset reduces projected claims, the bias is conservative against the buyer — it makes the saving look larger, never the cost. v2.1 discloses this rather than silently correcting it; the planned fix is a v1.1 exposure haircut for cross-marker correlation, which would lower the offset and is therefore deferred to its own version bump and re-attestation.

5. Cohort floor.

   Because the metabolic offset is drawn from the v1.0 forecast, the calibration inherits v1.0's minimum-cohort requirement (≥ 15 enrolled lives). Below the floor the procedure refuses to generate rather than return a non-credible figure.

6. One period, one forward year.

   The procedure calibrates a single connected period to a single forward year. It is not a multi-year reserve model and does not chain successive renewals.

§14. Version History and Change Control

v2.1 is additive to both v1.0 (unchanged, version-pinned) and v2.0 (which remains published and version-pinned). v2.1 changes exactly two things relative to v2.0: the §6 derivation of the credibility constant k (the value 400 and the worked numbers are unchanged) and the §9 band construction (single inherited band → two labeled bands from one variance model, adding the disclosed dispersion input τ). A change to any constant in the version block above, to the credibility constant, to τ's default, to the band construction, or to the offset-consumption rule constitutes a version bump and requires re-attestation. Attestations are version-pinned: an attestation against v1.0 or v2.0 does not transfer to v2.1, and a signed Workforce Claims Forecast records the methodology version it was generated under.

§15. Bibliography

Bühlmann, H. (1967). Experience rating and credibility. ASTIN Bulletin, 4(3). · Bühlmann, H., & Straub, E. (1970). Glaubwürdigkeit für Schadensätze. Bulletin of the Swiss Association of Actuaries. · Longley-Cook, L. H. (1962). An introduction to credibility theory. PCAS (limited-fluctuation full-credibility standards). · Herzog, T. N. Introduction to Credibility Theory. · Klugman, S., Panjer, H., & Willmot, G. Loss Models: From Data to Decisions. · Actuarial Standards Board. ASOP No. 23 (Data Quality), No. 25 (Credibility Procedures), No. 41 (Actuarial Communications), No. 56 (Modeling). · Society of Actuaries, group health renewal and manual-rating practice notes. · The metabolic-exposure derivation, per-marker coefficients, and bootstrap confidence-interval construction consumed as the offset M are specified in full in Methodology v1.0.