Too many candidates
Testing variables one at a time makes it hard to see which combination adds durable signal.
Risk modelling
Build superior models 10x faster.
Exakt fits and compares multiple Explainable Boosting Machines (EBMs) at once, using smart variable selection. Automatic smoothing greatly reduces the manual work between fits.
Modelling challenges
Manual work slows model development.
Manual smoothing
Repeatedly reshaping noisy functions takes time and can make decisions difficult to reproduce.
Lost model history
When alternatives overwrite each other, reviewers cannot see why one model was selected.
How Exakt helps
Build and compare models with less manual work.
EBM's are additive in nature and can serve as a drop-in replacement for classic GLMs with near-GBM performance. Exakt's optimised implementation trains a single EBM up to 6x faster and achieves up to 10% higher predictive performance than the base implementation.
Inspect every factor and smooth noisy effects
Explainable Boosting Machine (EBM) is additive, so you can inspect each factor's fitted effect on risk. Automatic smoothing reduces noise where data is thin and cuts the manual work between fits.
VariableDriver ageConstraintIncreasing
Original and auto-smoothed driver age relativities with exposure.
Choose the variables that improve your model
Compare candidate variable sets on holdout performance, calibration, stability and complexity. Use the gains from each addition to decide whether a larger model is justified.
Gini by model sizeHoldout · higher is better
Find the interactions worth keeping
Exakt automatically detects interactions worth testing. See which combinations matter and keep only those that improve the model.
InteractionDriver age × fuel typeHeightExposureColourInteraction factor
The illustrative interaction spans complete driver-age bands from 17 to 93 and over. Exposure peaks in the middle age bands and falls in the tails. The interaction factor ranges from 0.82 to 1.18, with 1.00 as the neutral point.
Three-dimensional driver age and fuel type interaction view.
Drag to rotate · Scroll or pinch to zoom · Hover or tap a bar for values
Control geographic smoothing to suit your model
Reduce noise in areas with less data while preserving the local patterns that matter.
CountryUnited KingdomViewPoint map
Loading geographic dataDrag to explore · Scroll to zoom
Lower local signalHigher local signal
SizeData volume
Interactive geographic smoothing preview
Compare candidates without losing model history
Fit multiple candidates at once and compare them on the same holdout data. Branch from any candidate to test variables, interactions or geography.
ValidationHoldoutRank byDeviance
Candidate evidenceSame holdout set
| Candidate | Deviance ↓ | Calibration → 1 | Stability ↑ | Terms | Decision |
|---|---|---|---|---|---|
| Parent v18 | 0.421 Baseline | 1.04 | 0.88 | 32 | Keep |
| Branch A | 0.408 −3.1% | 1.02 | 0.85 | 47 | Review |
| Branch B | 0.411 −2.4% | 1.00 | 0.93 | 26 | Candidate |
Selected for portfolio testBranch B reaches 1.00 calibration and the highest stability with 6 fewer terms than the parent. Branch A has lower deviance, but uses 21 more terms and is less stable.
Each addition creates a branch with its order, settings, diagnostics and decision recorded. Compare alternatives without overwriting the parent model.
Commercial decision
Use the risk estimate to find where you can compete.
A clearer estimate of expected loss shows where the current price leaves room for a better offer and where the insurer needs more premium, different terms, or less exposure.
Check the expected close rate and margin across the portfolio.
Change the price, terms, or exposure before the claims result arrives.