Actuaries have built this table for a century under other names: the experience frame behind a rate review. Machine learning did not change its shape, only how much can be learned from it. The columns are the application answers and policy facts (features), a unit of exposure such as policy-years or trip-weeks, and the outcome: number of claims, total paid and reserved, and the premium charged for comparison. An as-of date lets the model be validated on a later period than it was trained on, which is the only validation that means anything in insurance.
The quality of a training dataset is decided before any modelling: whether the features were captured as typed fields, whether exposure can be computed from dated policy changes, whether claims join to policies by key, and whether statuses mean one thing. A frame built from documents is expensive and noisy; a frame queried from a structured book is a day's work.
On Soldek
On Soldek the frame is a join across the exported tables: policies (exposure, premium) to quotes (answers per field, cover, channel) to claims (count, paid, reserve), with the product's field definitions giving each column its type. The full export from Settings → Your data contains all of them; the Client API serves products and submissions as JSON for a scheduled pipeline.
Your dataQuestions
- How large does an insurance training dataset need to be?
- For a frequency model, a few thousand policy-years with a few hundred claims already separates risk meaningfully; severity needs more claims and is usually modelled with fewer features. Structure matters more than size: a small clean frame beats a large reconstructed one.
- Should the training dataset include the premium?
- Yes, as a column, not as a feature. It lets you compute loss ratios per segment and compare the model's relativities with the tariff that priced the book, which is normally the point of the exercise.