Before an MGA commissions a pricing model, a data scientist or an insurtech partner will ask to see the data. The checklist below is what they are looking for. Score honestly: every 'no' is a step that will have to be done by hand before any modelling starts.
Capture
- Is every application question a named field with a type (number, date, option, yes/no), rather than free text?
- Are the answers stored on the quote record itself, not only in the generated document?
- Do the same facts have the same field name across products (the age of the insured is always the same field, not three spellings)?
- Are option lists fixed (a breed size is one of four values), not typed by the broker?
- Is the sum insured, the cover or package and the duration stored as data on every quote?
Pricing
- Is the premium computed from rate rows that exist as data, rather than from a spreadsheet beside the system?
- Can a premium from last year be re-computed from the stored answers and the rate rows?
- Are rate changes logged with who changed what and when?
- Where the underwriter overrides the computed premium, is the override visible as a difference between two numbers?
Decisions
- Are accept, refer and decline rules data (conditions on fields), not knowledge in someone's head?
- Is the underwriter's final action (quoted, declined, bound) a status with a timestamp?
- When a quote is declined, is the reason recorded?
Outcomes
- Does every claim point to its policy by key, in the same system?
- Are date of loss, cause, reserve, paid and settled stored as separate fields?
- Do claim statuses follow fixed transitions, so 'settled' means the same thing in every row?
- Are endorsements, cancellations and renewals recorded with effective dates, so exposure per row can be computed?
Access
- Can you export every table yourself, today, without a request to the vendor?
- Is the export a table (Excel, CSV, JSON), not a report or a PDF?
- Is there an API for the same records, so a pipeline can pull them on a schedule?
- Does the contract state that the data is yours, is never used to train anyone else's model, and leaves with you?
Scoring
Eighteen or more: you can start modelling on what you have. Twelve to seventeen: a focused clean-up, usually of the application answers, then modelling. Under twelve: the book cannot train a model in its current form, and the decision to make is about the platform, not about the algorithm.