ACORD 2026: 3 Claim Inputs That Reshape P&C Models

The $75 monthly cost of a no-code ML platform is less than a single actuarial hour, yet it can run the disagreement-adjusted risk (DAR) scoring that ACORD 2026 now demands. That price point is why the standard's three new claim inputs—climate exposure, model disagreement, and adaptation terms—are not just data fields but a modeling paradigm shift. Carriers that treat them as add-ons will miss the interaction effects.

The interaction is where the money hides. A per-claim reconciliation step prevents double-counting when all three inputs are present, while a portfolio calibration cost for the adaptation term—borrowed from building-performance research—recoups itself in a single loss event. A per-policy data quality check is the price of admission for accurate reserves.

The NAIC's 2025 reserve accuracy study found that a significant share of P&C loss reserves were off by significant amounts. With ACORD 2026's interacting inputs, that error rate can be cut—but only if carriers invest in the per-claim validation that the standard implies. The $16,440 weekly wage of a top soccer player is a reminder that money is often spent on the wrong things.

The Mechanism

The Insurance Institute for Highway Safety's 2024 study is the mechanical anchor for ACORD 2026's crash-pulse field: delta-v, the change in velocity during impact, explains a significant portion of the variance in injury severity. The standard now requires the "crash pulse" field to carry delta-v and impact angle from vehicle sensors. In a gradient-boosting model, that field enters as a continuous severity predictor rather than a categorical collision code, and that distinction matters because the model can learn that delta-v interacts with impact angle nonlinearly. A high delta-v at a perpendicular angle is not the same injury as a high delta-v at a glancing angle; the schema finally lets the model see that difference.

The second input is a cost index, not a static lookup. ACORD 2026 references CCC's Market Valuation for parts pricing and Mitchell's Labor Rate Survey for labor rates, both updated daily. According to those index benchmarks, this approach cuts cost-estimation error compared with static tables. The mechanism is time: a repair estimate written in February 2026 should not be priced from a table built months earlier. The reserve model needs to treat the index as a time-varying covariate, so a claims transaction timestamp aligns parts and labor prices to the actual repair date rather than the policy inception date.

The third input is behavioral. ACORD defines a composite claimant engagement score from claim-app login frequency, response latency, and document-upload timeliness. According to a State Farm pilot, low engagement scores correlate with a higher likelihood of fraudulent claims. The mechanism is not moral judgment; it is abandonment. Claimants who do not intend to pursue a valid claim stop engaging, and the model needs that signal to avoid over-reserving for claims that will close unpaid.

All three inputs are mandatory fields in the ACORD 2026 claims transaction schema, reversing their optional status in 2025. Treating that as a compliance task—"just update the XML"—misses the predictive shift. These are not administrative fields; they are severity, cost, and fraud covariates. But mandatory status also exposes the infrastructure gap. Crash pulse arrives as high-frequency telematics data, and the repair indices refresh daily, so the upstream pipeline must stream events and update covariates continuously. Legacy systems built around nightly batch extraction typically lack this capability; they either drop the telematics signal or aggregate it into a one-number-per-claim summary that destroys the delta-v timing signal.

None of this works if the model starts from ACORD's default priors. Those priors are population averages estimated from broad industry experience; your portfolio has its own distribution of vehicle models, repair networks, and claimant behavior. The mechanism—gradient boosting on crash pulse, live repair indices, and engagement scores—converges to the reduction in reserve error only after you run a validation study on your own historical claims and replace the defaults with portfolio-specific priors. That recalibration is not a tuning step; it is the step.

The Evidence

The evidence base for the ACORD 2026 claim inputs is stronger than the typical vendor white paper, but it carries a specific, quantifiable caveat: the gains are real, yet they are conditional on portfolio-specific recalibration. The most rigorous public evidence comes from a 2025 Casualty Actuarial Society working paper by Dr. Emily Chen, which analyzed auto claims and found that adding telematics impact severity to loss reserving models reduced mean absolute percentage error. This is not a marginal tweak; it is a material shift in reserve accuracy. However, the study's sample was drawn from a single telematics program with uniform driving behavior data, which means the prior distributions for impact severity were already well-calibrated to that specific portfolio. An insurer with a different geographic or demographic mix cannot assume the same gain without first testing the input against its own historical claims.

Operational evidence from the supply-chain side reinforces the thesis. According to a 2026 report by Deloitte's Insurance Analytics group, based on carrier implementations, insurers using real-time repair cost indices saw a reduction in claim cycle time and an improvement in settlement accuracy. The mechanism here is straightforward: repair cost indices replace static, periodically-updated labor and parts tables with live market data, which reduces the need for re-inspection and supplemental payments. But the settlement accuracy gain is the figure that matters for reserving—it implies that the ultimate claim cost is being pinned down earlier in the lifecycle, which directly reduces the variance in incurred-but-not-reported (IBNR) estimates. The cycle time reduction is a cash-flow benefit, but the accuracy improvement is what feeds the reserve error reduction.

The third input, claimant engagement scores, addresses a different failure mode: fraud leakage. A 2025 study in the Journal of Risk and Insurance showed that adding claimant engagement scores to traditional logistic regression models improved fraud detection AUC. This is a significant lift in discriminative power. The behavioral signal—how quickly a claimant responds, whether they provide consistent narratives, their interaction patterns with the claims portal—is a proxy for claim legitimacy that static policy data cannot capture. For loss reserving, the benefit is indirect but powerful: better fraud detection means fewer inflated payouts that distort the tail of the loss distribution. The AUC improvement translates to a meaningful reduction in the probability of a fraudulent claim being paid as legitimate, which is exactly the kind of tail risk that creates reserve volatility.

The aggregate evidence comes from the National Association of Insurance Commissioners (NAIC) 2026 data call, which revealed that early adopters of the ACORD 2026 standard reduced their loss reserve volatility year-over-year. This is the headline figure, but it must be read with the same caveat that applies to Dr. Chen's paper: the NAIC data reflects early adopters, who are self-selected for having both the data infrastructure and the actuarial sophistication to integrate these inputs properly. They are not a random sample of the industry. The reduction is the prize, but it is not the baseline expectation.

The critical bottleneck is infrastructure. According to a 2025 survey by Willis Towers Watson, only a minority of P&C insurers have the data infrastructure to ingest these inputs at scale. This is the binding constraint. The rest are not facing a modeling problem; they are facing a data engineering problem. Telematics impact severity requires streaming or batch ingestion of telemetry data, repair cost indices require API connections to parts and labor databases, and claimant engagement scores require a digital claims platform that logs every interaction. Without these pipelines, the gradient-boosting model is irrelevant.

The practical takeaway is that the evidence supports adoption, but it does not support blind adoption. The volatility reduction is the ceiling for a well-prepared insurer, not the average outcome. The validation study is not a regulatory hurdle; it is the mechanism by which the default priors in the ACORD 2026 schema—which are calibrated to a hypothetical average portfolio—are replaced with priors that reflect your actual claims history. Dr. Chen's gain was achieved because the telematics priors matched the sample. Your gain will depend on how well you can replicate that alignment with your own data. The infrastructure-readiness figure from Willis Towers Watson is the real gatekeeper: if you are in the majority without ready infrastructure, the modeling question is moot until you fix the data pipelines.

The Decision Framework

The decision of which ACORD 2026 input to prioritize is not a data-science question; it is a portfolio-stratification question. The three inputs—telematics impact severity, repair cost indices, and claimant engagement scores—do not compete on a single axis. They compete on five, and the optimal choice shifts depending on the composition of your book. The table below compares them directly, using the criteria that matter for a reserving actuary who has to defend the model to a state regulator and to an IT budget committee.

The practical sequence is this: run a validation study on your own historical claims to set portfolio-specific priors, as the canonical decision rule requires. If your auto book is telematics-heavy, prioritize the severity input and accept the streaming infrastructure cost. Otherwise, start with the repair cost index—it is the only input that gives you a meaningful lift without forcing you to defend a black box to a regulator. The engagement score should be a secondary feature, added only after the other two are stable and only if your claims data shows that app usage is not correlated with claim severity.

The three ACORD 2026 inputs are not neutral measurements; they are artifacts of the systems that produce them. Before you trust a gradient-boosting model that consumes them, you need to know where each one lies to you.

Telematics impact severity is a newer-vehicle bias in disguise. The crash-pulse field is derived from onboard sensors, and older vehicles simply do not have them. If your book skews older than the national average—say, a commercial auto portfolio of older work vans—the telematics severity input will be computed from a sliver of your actual claims. The model will appear to have strong predictive power on that sliver, but it is silent on the majority of your losses. The selection bias overstates the input's value precisely for the fleets that need it most.

Repair cost indices are not stable; they are volatile during supply-chain disruptions. According to a 2025 study by the Federal Reserve Bank of Chicago, parts price indices can swing sharply within a single month. A reserve estimate built on a June index may be stale by July. This is not noise to be averaged away; it is a structural feature of a globalized parts market. If your portfolio is heavy in makes with long-tail parts sourcing, the index is a lagging indicator that will systematically misprice your reserves during disruption windows.

What the Data Doesn't Tell You

Claimant engagement scores are confounded by demographics. A 2024 paper in the Journal of Behavioral Economics found that older claimants are less likely to use mobile apps. An older policyholder with a legitimate claim and a low engagement score is not a fraud signal; they are a channel preference. If your book has an older demographic profile, the engagement input will artificially depress scores across the board, and your model will learn to treat age as a proxy for fraud—a classic confounding trap.

The default priors are a national average, and averages lie regionally. The ACORD 2026 standard's default weights are calibrated on a national mix of repair costs and driving behavior. They fail to account for the difference between, say, urban New Jersey (high parts markup, dense traffic) and rural Montana (low markup, high-speed single-vehicle crashes). Using the default priors without recalibration will produce systematic over-reserving in one state and under-reserving in another, purely as a function of geography.

The headline improvement is not uniform. According to a 2026 RAND Corporation analysis, the improvement in reserve accuracy reported by early adopters varies across carriers. That means a meaningful number of carriers see no improvement, and some see degradation. The average is real, but it is not your destiny; it is a distribution you land in based on your portfolio's fit to the default priors.

The three inputs are not independent; they are collinear. Low engagement scores often coincide with high telematics severity—the same driver who is hard to reach is often the one driving aggressively. A naive additive model double-counts this shared signal, inflating the apparent effect of each input. You must test for multicollinearity before integration, or your gradient-boosting model will assign spurious importance to a single underlying factor.

The rule holds—adopt the inputs—but only after a validation study on your own historical claims. The improvement is a conditional promise, not a guarantee. It is earned only by the act of recalibration. Without it, you are not adopting a standard; you are importing a bias you do not yet understand.

Golden State Mutual is a hypothetical mid-sized auto insurer with a large auto policy count in California, and it is the cleanest worked case for the article's thesis because the carrier did one thing correctly: it treated ACORD 2026's default priors as a starting point, not an answer. The full set of before-and-after figures is in the table; the mechanism behind those figures is what follows.

The Shapley value split for this specific book is not a law. Golden State's telematics program already covered a substantial share of vehicles, so the telematics impact severity input was built from a signal the carrier owned before the standard arrived. The repair cost indices from Mitchell corrected a real-time pricing problem that was inflating physical damage case reserves. The claimant engagement score from the mobile app contributed least because app usage is not randomly distributed across claimants — a selection effect the model had to learn before it could trust the score. Those three contributions would change for a carrier with a different mix of telematics penetration, claims distribution, and customer adoption.

Choosing among the three ACORD 2026 inputs is not a data-science exercise; it is a portfolio-readiness audit. The reserve-error reduction cited in the thesis is an upper bound achieved under ideal conditions—a portfolio where all three inputs are available, clean, and regulator-accepted. Your portfolio is not that portfolio. The decision tree below is built from the constraints that actually determine whether gradient boosting will find signal or just amplify noise.

A Worked Case

Rule 1: Telematics penetration is your gating variable. If only a small share of your claims have telematics data, the impact-severity input will be a sparse column that the model either ignores or, worse, uses to partition a tiny, unrepresentative subsample. In that case, do not invest in telematics impact severity. Make repair cost indices your primary input instead—they are available for every claim with a parts or labor line item, which in most auto books is effectively all physical-damage claims. The mechanism here is straightforward: a gradient-boosting model can exploit interactions only when the interacting columns have sufficient non-null overlap. Sparse telematics data forces the model to send most claims down a branch that never sees the telematics feature, effectively wasting the tree depth on a feature that applies to a minority of rows.

Rule 2: Run a validation study on your own historical claims before committing to any full implementation. The ACORD 2026 standard's default priors were estimated on a pooled multi-carrier sample, and those priors encode the average relationship between the inputs and ultimate loss emergence. Your book has different geography, different policy language, different claims-handling workflows. The validation study needs a large sample of claims with complete data on all three inputs plus a mature loss-development triangle so you can measure predictive lift against actual emergence. The methodology proposed in the European building-performance prediction literature—computing an adaptation term that reconciles numerical predictions with standard predictions—is directly transferable here: you are estimating the difference between the ACORD default prior and your portfolio's true conditional relationship, then recalibrating. Without that adaptation term, you are importing someone else's actuarial assumptions and calling them your own.

Rule 3: Check your state regulator's position on repair cost indices before building anything around them. The NAIC's 2026 rate filing guidelines are the relevant document. Some states have signaled acceptance of third-party repair cost indices as a legitimate basis for loss-development assumptions; others have not, and a handful have explicitly questioned whether a vendor-supplied index constitutes a permissible ratemaking input versus an unverifiable external data source. If your regulator does not accept repair cost indices in rate filings, deprioritize that input entirely and focus on telematics impact severity or engagement scores—whichever your portfolio can support. There is no point engineering a feature that your filed rates cannot legally use, because the reserve-error reduction only matters if it flows through to your booked reserves, which regulators audit.

How to Choose Well

Rule 4: Engagement scores are the most behaviorally loaded input, and they carry the most hidden bias. The score is constructed from policyholder interactions—app usage, claim-filing behavior, response times to adjuster communications. If you collect that data without an opt-in consent mechanism, you are not only exposing yourself to privacy risk under state law, but you are also building a feature that systematically excludes older policyholders and those with low digital literacy. The fix is twofold. First, ensure the collection process is privacy-compliant: opt-in consent, clear disclosure of what is tracked, and a mechanism for withdrawal. Second, include age and digital literacy as control variables in the model. The gradient-boosting framework handles this naturally—it will find the interaction between engagement score and age if you let it—but only if you feed it the controls. Without them, the engagement score will proxy for demographics rather than for claims behavior, and your reserve estimates will inherit that confounding.

Rule 5: Use a gradient-boosting or random forest model, not a linear model. The three inputs are correlated by construction: a severe impact (telematics) drives higher repair costs (repair cost indices), which in turn may drive a more engaged claimant (engagement scores). A linear model will double-count these correlated effects because it cannot represent the conditional structure—it will add the marginal contribution of each input as if the others were held fixed, when in fact they move together. Gradient boosting and random forests handle this by construction: they partition the feature space, so the model can learn t

Frequently Asked Questions

How can a small carrier run the three-model DAR scoring without hiring more actuaries?

The $75 monthly cost of a no-code ML platform is less than a single actuarial hour, yet it can run the disagreement-adjusted risk (DAR) scoring that ACORD 2026 now demands.

Why does ACORD 2026 require raw delta-v and impact angle instead of collision codes?

In a gradient-boosting model, the crash-pulse field enters as a continuous severity predictor so the model can learn that delta-v interacts with impact angle nonlinearly, making a high delta-v at a perpendicular angle different from a high delta-v at a glancing angle.

When all three mandatory inputs are present on a claim, what prevents double-counting?

A per-claim reconciliation step prevents double-counting when all three inputs are present.

What exactly is in the repair cost index that ACORD 2026 uses?

ACORD 2026 references CCC's Market Valuation for parts pricing and Mitchell's Labor Rate Survey for labor rates, both updated daily.

What did the NAIC 2026 data call show about early adopters of ACORD 2026?

The NAIC 2026 data call revealed that early adopters of the ACORD 2026 standard reduced their loss reserve volatility year-over-year, though they were self-selected for having the data infrastructure and actuarial sophistication to integrate these inputs properly.

Why can't I just keep ACORD's default priors instead of recalibrating to my own portfolio?

ACORD's default priors are population averages estimated from broad industry experience, so you must run a validation study on your own historical claims and replace the defaults with portfolio-specific priors.

Quick answers

Sources: Reddit, arXiv, arXiv, arXiv, Reddit

Also worth reading: Analyzing the true impact of inflation on insurance claims reserves: Analyzing the true impact of · How AI Is Reshaping Actuarial Consulting in 2026: How AI Is Reshaping Actuarial · How AI Insurance Checkers Are Transforming Coverage in 2026: Real-Time Risk Insights: How AI Insurance Checkers Are