Skip to main content
    Kriv AI

    Portfolio Accelerator · Insurance · Databricks

    Actuarial Reserving Software Design on Databricks: An Implementation Blueprint for Insurance Carriers

    This is a design for actuarial reserving software on Databricks: a documented architecture for ML-augmented loss development triangles and reserve scenario modeling, governed by Unity Catalog.

    problem

    The Reserving Problem: Why Loss Development Is Still Spreadsheet-Bound

    Loss development triangles and reserve estimation still run in spreadsheets at most carriers, even though the underlying chain-ladder, Bornhuetter-Ferguson, and Mack methods have well-established open-source implementations. A Databricks-native carrier wants those methods running on the same lakehouse already holding its claims data, with Unity Catalog governance and audit trail built in, not a separate actuarial tool bolted on the side.

    demo

    Architecture: How the Databricks Actuarial Reserving Accelerator Is Designed

    This page shows the architecture for Kriv AI's actuarial reserving accelerator on Databricks. In the interest of transparency: the reference-research phase is complete; the platform build, reserving engine, and MLflow integration are specified but not yet built. Build-Demo-Vanish means we build once, demonstrate the working pattern, then retire the accelerator, and right now this one is still at the design stage, which we'd rather say plainly than imply otherwise.

    Unity Catalog, Lakehouse, and Governed Actuarial Data Pipelines

    The design calls for a medallion architecture (bronze, silver, gold) on Delta Lake, with Unity Catalog governing catalog, schema, and table access for actuarial data, and Delta Lake time travel intended to provide the reproducible audit trail SOX and regulatory reviews require.

    ML-Augmented Loss Development Triangles and Scenario Modeling

    The reserving engine is designed around the well-established chainladder-python methods, volume-weighted and simple-average Chain Ladder, Bornhuetter-Ferguson with expected loss ratio inputs, Cape Cod with earned-premium weighting, the Mack stochastic model for process and parameter variance, and Bootstrap ODP for a full predictive reserve distribution, with a Bayesian reserving layer using PyMC as a planned extension for uncertainty quantification.

    exact query answer

    What is actuarial reserving software?

    Actuarial reserving software is a tool that helps actuaries estimate what an insurer should hold for claims that have already happened but are not fully paid. Most tools develop loss triangles with established methods such as chain ladder, Bornhuetter-Ferguson and the Mack model, then let the actuary select assumptions and document them. Our own actuarial reserving accelerator is a design for Databricks, not shipped software: the reference research is complete, and the platform build, reserving engine and MLflow integration are specified but not yet built. This page explains the design so a carrier can judge whether it fits.

    exact query answer

    What is an actuarial reserving solution?

    An actuarial reserving solution is the whole setup around the calculation: the claims data pipeline, the reserving methods, the rules on who can see and change what, and the record that lets a past estimate be reproduced. Our design places that on Databricks. Claims data moves through bronze, silver and gold Delta tables, and Unity Catalog, which Databricks describes as its unified governance layer for data and AI, enforces permissions, records lineage and logs activity. The actuary stays in charge of method selection and assumptions. This design is not yet built, so ask any vendor, including us, to show a working run on synthetic data before you commit.

    exact query answer

    What is AI loss reserving software?

    AI loss reserving software adds machine learning to established reserving methods instead of replacing them. In our design, ML augments the loss development triangles and scenario modeling, while chain ladder, Bornhuetter-Ferguson, Cape Cod, Mack and Bootstrap ODP stay as the methods the actuary can compare against. A Bayesian layer using PyMC is a planned extension for uncertainty. The reserve stays a professional judgment, so the software suggests and documents and an actuary decides and signs. None of this is built yet. We say so because AI reserving claims are easy to make and hard to check.

    exact query answer

    What is actuarial reserving?

    Actuarial reserving is the work of estimating how much an insurer should hold for claims it owes but has not yet paid. That money is known as a loss reserve, and it includes an allowance for claims that have already happened but have not yet been reported to the insurer. Reserving applies statistical methods to claims data, and the result is an estimate that an actuary reviews and documents. The design on this page covers the data and method layer of that work. The judgment stays with the actuary.

    deepening

    Can a reserve estimate be reproduced on Databricks?

    A reserve estimate can be reproduced on Databricks only if the data behind it can be read back as it stood on the day of the run. Databricks documents Delta table history, which returns the operations, user and timestamp for each write to a table, and time travel, which queries an earlier version of a table. It also documents a limit: in Databricks Runtime 18.0 and above, time travel queries are blocked for versions older than the table's deletedFileRetentionDuration setting, which defaults to 7 days. So the design would set retention to match the carrier's own record rules, or save each quarter-end reserving run as its own dated table. This is design intent, not a running system.

    deepening

    How would MLflow and lineage record a reserving run?

    In the design, each reserving run would be logged with MLflow tracking, which Databricks describes as a way to log parameters, metrics, tags and artifacts and to compare runs inside an experiment. For reserving, the method, the assumptions the actuary chose and the resulting estimates would sit together in one record. Unity Catalog lineage adds the other half: Databricks says data lineage shows where data came from and where it goes, including which queries and files populate a table. A reviewer could then trace a reserve figure back to the claims tables behind it. The MLflow integration is specified and not yet built.

    exact query answer

    What is insurance reserving software?

    Insurance reserving software is a tool that helps an insurer estimate the money it needs to hold for claims that have occurred but are not fully paid. That money is known as a loss reserve, and it is reported as a liability in the insurer's financial statements. Software for this work usually covers the claims data, the reserving methods and the record of the assumptions chosen, and an actuary reviews the result. The design on this page is for carriers that run Databricks. It is a design and not yet built, so it is not software you can license today.

    status

    What's Real Today, and What's Roadmap

    Honestly: the reference-research phase is complete, 21 repositories reviewed across actuarial reserving methods, statistical modeling, Databricks infrastructure, and MLOps tooling. The platform build, PySpark reserving implementation, MLflow experiment tracking, and compliance reporting described here are all specified but not yet built and exercised on synthetic claims data.

    extension

    From Reserving to Underwriting: Extending the Databricks Pattern Across the Policy Lifecycle

    The same Unity Catalog governance and Delta Lake medallion pattern designed for reserving is meant to extend across the policy lifecycle, from underwriting risk scoring through claims and reserving, so a Databricks-native carrier gets one governed data platform instead of a separate tool per function.

    compliance

    Regulatory-Ready by Design: NAIC Model Bulletin and Audit Trail Requirements

    The architecture is designed to support ASOP 43 documentation of methods, assumptions, and reasonable-estimate ranges, with Unity Catalog access controls and Delta Lake time travel providing the audit trail SOX Section 404 and NAIC Model Bulletin reviews expect from a production reserving model.

    differentiation

    Why Kriv AI for Your Databricks Work

    Build-Demo-Vanish: Our Methodology for Starting From a Written Design

    Kriv AI's Build-Demo-Vanish methodology starts from a written Databricks design and reserving methods that are already selected and documented, so a scoped engagement can begin from a written design instead of a blank page.

    Straight answers

    Frequently asked questions about Actuarial Reserving Software Design on Databricks: An Implementation Blueprint for Insurance Carriers

    Is the actuarial reserving accelerator built and running on Databricks today?

    No, and we want to be direct about it. The reference-research phase is complete (21 repositories reviewed), but the platform build, reserving engine, and MLflow integration are specified but not yet built and exercised on synthetic claims data.

    What reserving methods does the design use?

    The well-established chainladder-python methods: volume-weighted and simple-average Chain Ladder, Bornhuetter-Ferguson, Cape Cod, the Mack stochastic model, and Bootstrap ODP, with a planned Bayesian layer using PyMC for uncertainty quantification.

    What role does Unity Catalog play?

    Unity Catalog is designed to govern catalog, schema, and table-level access for actuarial data, with Delta Lake time travel providing the reproducible audit trail SOX and regulatory reviews require.

    Is any real claims data used?

    No. This accelerator, once built, will run entirely on synthetic claims data engineered to mirror real loss-triangle structures. No real policyholder or claims data is used.

    What does Build-Demo-Vanish mean?

    It's Kriv AI's methodology: build a working proof of concept on synthetic data and real infrastructure, demonstrate the pattern to a prospect, then retire the accelerator once a real, bespoke engagement is scoped.

    Can Kriv AI build this out for our carrier's Databricks workspace?

    Yes. A scoped engagement can build this architecture against your real Databricks workspace and claims data. Contact us to discuss scope and timeline.

    Does actuarial reserving software replace the actuary?

    No. In our design the software develops loss triangles and lets the actuary compare methods, and the actuary selects the method and assumptions, reviews the result and signs it. Be cautious with any tool that shows a reserve as a finished answer without the assumptions behind it.

    What should we ask an actuarial reserving software vendor to show us?

    Ask for a working run on synthetic claims data, the list of methods it supports, how each assumption is recorded and how a past estimate can be reproduced. We apply the same test to our own design. It is not yet built, so we cannot show a run today.

    Can a past reserve estimate be reproduced on Databricks?

    It can if the data is kept. Databricks documents table history and time travel, and in Databricks Runtime 18.0 and above time travel is blocked for versions older than the deletedFileRetentionDuration setting, which defaults to 7 days. A design has to set retention or save dated snapshots of each reserving run.

    What does Unity Catalog record about a reserving workload?

    Databricks says Unity Catalog enforces access control, tracks lineage as data and AI assets are used and logs activity for auditing. For a reserving workload that would show who read the claims tables and which tables feed which output. How long logs are kept and who reviews them are decisions for the carrier.

    What are the benefits of automating the actuarial reserving process?

    Automating the actuarial reserving process means running loss triangles and reserving methods on one platform instead of by hand. The benefits our design aims for are a repeatable run, the assumptions recorded next to the estimate, and lineage, which is a record of which claims tables a figure came from, using Databricks table history, MLflow tracking and Unity Catalog lineage. We have not measured time savings and do not claim any, because the design is not yet built. The actuary still selects the methods and reviews the result.

    Ready to talk through the design for your data model?

    Bring your requirements to a working session and we'll walk through the design.

    Book a Discovery Call