Building insurance pricing models from US rate filings

In the US, most property and casualty insurers must file their rates and rating rules with each state's insurance department before using them. Those filings are generally public records. They contain base rates, rating factors, rules and, often, worked premium examples. Together they make a filing one of the most detailed sources of pricing information available anywhere.

With the right process, a filing can be turned into a working pricing model. AI now makes that process much faster than it used to be.

What a rate filing contains

Filings vary by state, line of business and insurer, but a rate or rule filing usually includes some or all of the following:

• Rate pages. Base rates and the tables of rating factors, such as territory, age, construction class, deductible and limit factors.
• A rules manual. Eligibility rules, how factors combine, rounding, minimum premiums, fees and surcharges.
• An actuarial memorandum. The insurer's explanation of the rate change and the analysis behind it, including the proposed rate level change and how it was derived.
• Supporting exhibits. Loss experience, trend and development, expense provisions and the derivation of factors.
• Premium examples. Sample calculations showing how a premium is built for specific risks. These are especially useful, because they show the order of operations and give exact results to test against.
• Correspondence. Questions from the regulator and the insurer's responses, which sometimes lead to changes in the rates that were finally approved.

Many filings also refer to a rating bureau, such as ISO, NCCI or AAIS, and adopt the bureau's loss costs or rules rather than restating them.

Where filings come from

Most states accept filings through SERFF, the System for Electronic Rate and Form Filing operated by the NAIC. Many states make filings publicly searchable through SERFF Filing Access or their own department of insurance website. Coverage, search tools and the documents released vary from state to state, and some exhibits may be withheld as confidential.

Turning a filing into a model

The work follows the same steps as building a model from any rating guide, with some extra steps specific to filings.

1. Find the full chain. A single filing often changes only part of a rating plan. To build a complete model, you need the filing that set up the programme and every later filing that changed it, up to the version in force.
2. Extract the rate tables. Pull every table from the rate pages into structured data, keeping the exact values and band boundaries.
3. Reconstruct the algorithm. Use the rules manual and premium examples to establish which factors apply, in what order, and how rounding, minimum premiums and fees work.
4. Add the rules. Encode eligibility rules as declines and referrals.
5. Test against the premium examples. Each worked example in the filing becomes a test. The model is complete when every example matches to the cent.
6. Record the source. Keep a note of which filing each table and rule came from, so any value can be traced back to its source document.

How AI helps

Filings are long and inconsistent. A single programme can run to hundreds of pages across several documents, with tables in scanned PDFs and rules spread over different sections. This is the kind of material AI handles well:

• Reading and extraction. AI can locate rate tables in PDFs and spreadsheets and convert them into structured data, including tables split across pages.
• Following the chain. AI can identify which earlier filings a filing refers to, and which pages each one replaced.
• Drafting the algorithm. From the rules manual and premium examples, AI can draft the order of calculation for a person to review.
• Checking. AI can run the premium examples as tests and explain any mismatch, for example a rounding step applied in a different place or a missing factor.

A pricing consultant reviews every table and rule against the filing and makes the decisions the filing leaves open. The tests built from the filing's own examples show whether the model is correct.

Limits to plan for

Filings are detailed, but they are not always complete. The main gaps are:

• Bureau materials. Where an insurer adopts bureau loss costs or rules by reference, the bureau's own manuals are usually not in the filing. They are copyrighted and normally require a licence from the bureau.
• Redacted content. Some states allow insurers to keep parts of a filing confidential, such as proprietary scoring models or detailed exhibits.
• External scores. Factors based on insurance scores, credit-based scores or third-party data often depend on models that are not included in the filing.
• Approved versus filed. Rates can change during review. Check the correspondence and the final disposition to make sure you are building what was approved.
• Effective dates and states. Each state has its own filing and effective date, and a programme can differ from state to state.

These gaps do not stop a model being built. They need to be identified early, and the missing pieces sourced or replaced with clearly documented assumptions.

Common uses

• Moving your own filed rates into a new rating engine. The filing is the approved version of your rates, so it is the right source to build and test against.
• Expanding to a new state. Starting from your existing filings and adjusting for state-specific rules and factors.
• MGA programmes. Building the rating for a programme written on a carrier's filed rates.
• Market benchmarking. Understanding how other insurers structure their rating, which factors they use and how their rate levels compare.

When using other insurers' filings, use them to understand the market and inform your own pricing. Your own rating plan should be based on your own analysis and filed in your own name, and copyrighted bureau material should be used only under licence.

Getting started with Ptally

Ptally builds pricing models from rating guides, spreadsheets and US rate filings. Our consultants use AI to extract the tables, reconstruct the algorithm and turn the filing's premium examples into a test suite, then launch the model on the Ptally rating engine. There is no implementation fee and nothing to pay upfront. Talk to our team about the filings you want to build from.