A 30-minute run-sheet for presenting the Risk & Rating Engine. It covers what it does, how the pieces connect, what to click, what to say, and how to handle the hard questions.
"RARE is GGI's own pricing workbench. Actuaries take real policy and claims data, build a statistical rating model, review and approve every factor, and publish it as a live rating API that quoting systems call in milliseconds. No spreadsheets, no hand-copied tariffs."
Motor tariffs live in spreadsheets and desktop tools. Getting a new rate into the quoting system is manual, slow and hard to audit.
One governed pipeline: data in → GLM analysis → reviewed rating factors → loadings → approved, versioned model → production API.
Faster rate changes, full audit trail of who approved what, and every channel (direct, broker, portal) quoting from the same approved model.
Walk the diagram left to right, top to bottom. Every box is a screen in the app and an API behind it.
Use the pre-built product 5525 (Private Comprehensive Leasing) or 5504 (Individual TPL) model for steps 5–10. Both have a solved severity GLM. Only create a new model live up to the data import, so the GLM fit can't fail on stage.
"Role-based access. Every model lives here with its status, from draft through approved and live."
"A model is the container. Data, GLM runs, factors and versions all hang off it, so the full history stays in one place."
"No exports or emailed files. RARE reads straight from GGI's policy and claims systems for the window you choose. Each import is a versioned snapshot, so a model can always be reproduced."
"The actuary decides which fields drive price. Fields with no spread, like a single branch or one vehicle make, are excluded so they don't add noise."
"Before modelling, we catch factors that tell the same story twice. Highly correlated pairs distort a GLM, so one side gets dropped here."
"This is the industry-standard actuarial method. On 5525, over-turning claims come out at about 6% of the act-of-god severity, with p = 0.000016, and that matches the raw claims almost exactly."
Severity · Gamma / Log · target ESTIMATED_NET_AMOUNT Act of god (reference) ≈ SAR 4,881.67 Over turning × 0.0618 p = 0.000016 AIC 171.59
"The statistics suggest, the actuary decides. Every factor can be smoothed, capped or overridden before it goes anywhere near a live quote, and each change is recorded."
"Once approved, a version can't be changed. A new rate means a new version. Registration decides which channel uses it and from when."
/api/rating/calculate with the API key"This is how OptimaX or a broker portal gets a price. The caller sends the risk, and RARE picks the registered model for that channel and product by itself."
{ "RiskPayload": { "NATURE_OF_LOSS": "Act of god" } }
→ Technical 1.00 · Loadings 0.4553 · Final 1.4553
{ "RiskPayload": { "NATURE_OF_LOSS": "Over turning" } }
→ Technical 0.10 · Loadings 0.0455 · Final 0.1455
Present these as relative rates. The base rate isn't multiplied in yet, so don't call them SAR premiums.
"You don't have to rebuild what already works. Your current tariff from other tools can go live through RARE's API on day one while new models are developed."
"Which Motor product should we take through first with full policy exposure? Then we can agree the rollout plan."
What the actuary does on each tab, from raw data to a live rating model, and when to move to the next tab. The tabs inside a model follow this order left to right.
Enter the name, line of business (Motor), product code, country, currency and owner. This only creates the container that everything else attaches to.
There's no cleaning screen in the app. If the profile shows bad data, fix it at the source or in the CSV and import a new version.
Move on when you have one merged dataset version with sensible row counts, a positive claim amount column and an exposure column.
Select the dataset version. Every column appears as a row, and for each one you set:
| Setting | What the actuary decides |
|---|---|
| Type | Continuous (age, sum insured, vehicle value) or Categorical (make, region, nature of loss) |
| Included | On for candidate rating factors; off for IDs, dates, policy numbers and personal data |
| Target | What is predicted: claim count for Frequency, claim amount for Severity |
| Exposure | Earned years or policy days (Frequency model only) |
| Offset | Rarely used; for a known fixed adjustment |
Then click Analyze on each candidate factor:
Move on when every included factor has zones or groups with reasonable volume in each level.
Pick the dataset version, set the threshold (for example 30%) and click Run correlations. The Cramér's V matrix highlights pairs above the threshold, such as vehicle value and vehicle make.
Actuary decision: keep the factor that is more predictive (from the one-way analysis) or easier to collect at quote time, and switch the other off in Variables.
| Frequency run | Severity run | |
|---|---|---|
| Component | Frequency | Amount |
| Distribution / link | Poisson / Log | Gamma / Log |
| Target | Claim count | Claim amount |
| Exposure | Exposure column | None |
Add the factors, choose a modelling form for each (Zones, Categorical, Linear, Polynomial or Spline), save the configuration and click Run. The job runs in the background. Then review:
Move on when both models are stable. Write down both GLM Run IDs.
Go to Versions → New version. The version must exist first, because premium, rating factors and loadings are all saved against it.
Add a version with the Frequency run ID, the Amount run ID and the correction term (1.0 by default), then click Run. The app writes one rating factor for each factor level.
For each level you see the Statistical value from the GLM. Set the Approved value and Final rating value: smooth uneven steps, round values, cap extreme ones, or override where business judgement disagrees. The app records who approved each change and when.
Add expense, commission, risk margin and profit, each as a percentage or a flat amount, in the order they should apply.
/api/rating/calculate uses it.Done: the model is live for that channel from its effective date.
For you only: current gaps. A factor's reference level can only be set through the API, not on screen. The GLM and Consolidated Premium screens ask for dataset, variable and run IDs as typed numbers, and the Variables table doesn't show variable IDs. Consolidated Premium and Approve don't warn when a GLM didn't fit properly.
It's GGI's own platform, with no per-seat licence, and it connects directly to GGI's databases and quoting systems. Existing tariffs from other tools can be imported, so it's a bridge, not a forced migration.
The GLMs run on statsmodels, a standard, peer-reviewed Python library used widely in actuarial and academic work. Every run stores its coefficients, p-values, AIC and deviance for review.
Nobody can change one in place. Approved versions are locked. A rate change is a new version that goes through approval and registration again, with users and roles controlled in Admin.
One API call. Factors are precomputed at approval time, so calculation is a lookup and multiply with no model fitting at quote time.
Yes. Registrations have effective dates, so you point the channel back at the previous approved version.
The pipeline is line-agnostic: model, data, GLM, factors, API. Motor is the first line configured.