Risk modelling, profit forecasting, and opportunity identification - we find where your money is lost, where it's made, and where the data is telling you something your team hasn't caught yet.
Clients anonymized.
The market floods with new car brands faster than statistics accumulate — often without adequate parts distribution, service infrastructure, or repair history. Brands appear and disappear before classical actuarial analysis becomes possible. High and heterogeneous repair cost inflation compounds the issue: it varies significantly across brands, and without data, estimation is impossible — the variance is too large. Standard pricing and reserving methods fail for a significant portion of the portfolio: insufficient data, and expert knowledge about relationships between brands, repair processes, and parts costs goes unused.
Instead of waiting for statistics on each brand separately, we built a model that uses structural similarity between car brands. Brands serviced at the same repair shops, with similar parts logistics, or comparable design characteristics are grouped — information from some brands enriches estimates for others. We used hierarchical Bayesian regressions where brands with single observations are pulled toward their similarity group, compensating for data scarcity. As new loss data arrives, estimates refine automatically.
Coverage of ~30 new mass-market brands within 6 months of market entry — without waiting for proprietary statistics. Identified non-obvious similarity clusters: several lesser-known brands ranked top in loss parameters, which wouldn't have been discovered without structural modeling. Enabled reserve estimate refinement as loss development data arrives, without waiting for period closure. Honest limitation: for unique vehicles without mass-market analogs, estimation is impossible — such objects are explicitly excluded from the model.
A specialized independent actuary needed a pricing model that could be updated incrementally as new policy data arrived — without full retraining. Standard approaches either couldn't handle the data quality issues typical of actuarial datasets, or required infrastructure that didn't make sense for a solo practice.
We built a GLM-based model with batch learning capabilities, specifically adapted to the noise patterns and missing data common in actuarial records. The model accumulates knowledge from new data without discarding historical calibration.
With a full historical dataset too large for a single-pass fit on standard hardware, the choice was between sampling — and losing information — or incremental fitting across the full history. We built batch-based IRLS so the model sees all the data without requiring it all at once.
The model accumulates evidence from new data without discarding historical calibration — update time dropped from days to hours. During high-volatility periods, when full retraining would have been unreliable, the incremental approach held.
A Mexican insurance startup was pricing crop coverage for corn using spreadsheets and generic national weather data. The model didn't account for localized weather patterns or seasonal timing — leading to mispriced risk and exposure to correlated losses that the generic data couldn't show.
We built a local climate downscaling model calibrated to historical weather patterns, with crop-specific exceedance thresholds. Combined with an extreme value model for tail events, this produced a full loss distribution for index insurance payouts.
The client moved from gut-feel pricing to a defensible, data-driven model — and discovered that Sinaloa's risk profile looked nothing like the generic national averages they'd been using. Pricing corrected accordingly before exposure materialized.
A reinsurance portfolio of 60 treaties looked compliant on paper — but the company had no systematic way to know which treaties actually transferred meaningful risk and which didn't. The standard 10-10 test measures probability of loss but ignores how severe those losses could be. Two treaties passing the same test can differ by 25x in real risk transferred. With only 5-15 large loss observations per treaty, estimates were unreliable, and there was no sensitivity analysis to distinguish robust conclusions from borderline cases that could fail under audit.
We built an actuarial framework for Expected Reinsurer Deficit (ERD) quantification across the entire portfolio, replacing point estimates with sensitivity-tested results. Large losses modelled as compound Poisson-Pareto using Peaks-Over-Threshold theory. Tail index estimated via fast MLE with Hill plot diagnostics. ERD computed via Monte Carlo simulation with analytical lower bounds for verification. Mandatory sensitivity testing across tail index, loss frequency, and discount rate for borderline treaties. Each treaty produces documented actuarial output suitable for governance and audit.
Portfolio-wide ERD range revealed 25x dispersion between highest and lowest values — commercially negotiated structures vary widely in actual risk transferred, independent of premium size. 2 of 60 treaties identified as structurally non-compliant under any consistent parameter set, now under commercial renegotiation. Portfolio assessment time reduced from one month to two days — a 15x efficiency gain — while producing sensitivity-tested actuarial opinions rather than point estimates. ERD findings now incorporated into renewal negotiations with objective, documented basis for treaty structure discussions.
Within a single year, litigation costs in one major insurance line grew 60% — driven not by external factors alone, but by changes in claims settlement decisions. Court settlements averaged 1.4x the original claim amount — before legal fees. 71% of disputed cases closed within the first year, leaving a significant tail of accumulated liabilities. Standard regression couldn't separate the causal effect of operational decisions from case characteristics — the decision (pay/deny) enters as just another feature, with no causal identification.
We formalized the settlement process as a directed acyclic graph (DAG): company decision → client reaction → court escalation → resolution time → final payout. Each node is a probabilistic model parameterized by case features and prior node states. A structural causal model (SCM) built on top enables do-calculus interventions — setting a specific company decision and estimating the counterfactual cost distribution. This allows causal inference: quantifying how outcomes would change under alternative operational decisions, not just correlations.
Quantified the expected financial impact of alternative settlement decisions by claim segment — giving the claims team a defensible, causally-identified basis for operational changes. Built a monitoring tool for tracking how changes in decision-making systems affect litigation cost dynamics — without waiting for full loss development. Model validation requires several years of development; results are prospective.
A pharmacy software platform had a regulatory compliance crisis hiding in plain sight. A mandatory medication — saline solution in a specific dosage — had become impossible to order due to a pricing ceiling and rising costs. Pharmacies were legally required to stock it. No one saw it coming because no alert existed for this kind of structural failure.
We modeled how procurement managers respond to shortage signals — and trained a change point detection algorithm to recognize the pattern early. The system identifies when a medication is trending toward unorderable before it becomes a compliance event.
Pharmacies avoided significant regulatory fines. The platform gained a competitive differentiator: predictive procurement alerts. The detection framework extended to other critical medications across the supply chain.
Feeling lost with technical decisions on data & machine learning? Want to predict losses, profits and identify risks based on data, but you don't know where to start? Get an expert perspective without committing to a full project.
Book a free intro call →Send the project details to xanaverte@axiolyze.com or
Book a call to discuss →Custom mathematical model development, data pipeline build & any data and machine learning solution your business needs. Delivered with full documentation and a handoff session.
Send the project details to xanaverte@axiolyze.com or
Book a call to discuss →
A no-code & modeling environment software built for actuarial workflows. Prototype models in an evening, not a month. Built-in regulatory alignment.
Xana Verte is a mathematician, data scientist, and actuarial data scientist. She came to insurance data through financial consulting — and then fell deep into actuarial statistics and modelling.
At VSK, one of Russia's largest insurers, she built pricing models for CASCO and other insurance lines, developed Bayesian causal models during volatile periods, and designed monitoring systems for portfolio cohort behavior. She's spent years working with datasets that break standard tools: sparse, noisy, small-sample, rare-event.
Her work on agroclimatic risk was presented in Malaysia. Her diagnostic framework for structural model misspecification is ongoing research. She's built models for clients in Mexico, India, and across Eastern Europe.
Axiolyze is the structure she founded around that work. Axiolyze is supported by a team of actuaries, mathematicians, researchers, software engineers and designers.
Yes. The core methods — change point detection, predictive modeling, risk quantification — apply wherever you have time-series data and consequential decisions. We've worked in pharmacy supply chain, agriculture, and financial modeling.
Axiolyze helps actuaries with mathematical and data solutions. Formal actuarial opinions and regulatory sign-off remain with credentialed actuaries.
It's actually where we're most useful. Most of our work involves datasets with missing values, irregular reporting, small samples, or data quality issues that standard tools handle poorly. We build for that.
It's a low-code/no-code actuarial modeling environment we're building — designed so actuaries can prototype models in an evening instead of a month. Join the waitlist and we'll keep you updated as it develops.