Why long-tail commercial lines strain the triangle
Most non-life reserving in India still runs on the chain-ladder method and its relatives, which project ultimate losses from an aggregate development triangle of paid or incurred amounts by accident and development period. The method is transparent and well understood, and for short-tail, high-volume lines it works well. The strain shows on long-tail commercial lines, liability, engineering, and other covers where claims take years to emerge and settle.
The difficulty is that aggregation throws away information. A triangle collapses every claim in a cohort into a single cell, so the heterogeneity within that cohort, a few large, slow liability claims sitting beside many small ones, disappears into an average development pattern. For lines where the tail is driven by a small number of large, individual claims, the aggregate pattern is a blunt instrument, and the IBNR provision, the reserve for claims incurred but not reported, is among the most material and most uncertain reserves a non-life insurer holds.
That is the opening for machine learning. The methods do not promise a single right answer where chain-ladder gives a wrong one; they offer a way to use the per-claim information that aggregation discards, which matters most exactly where the triangle is weakest.
Survival models for IBNR and individual-claim reserving
Two strands of method are worth separating, because they address different parts of the reserve.
The first is survival analysis for IBNR. Because IBNR is fundamentally about claims that have occurred but not yet been reported, the reporting delay is a time-to-event quantity, which is what survival analysis is built to model. Methods based on survival analysis have been developed to estimate IBNR claim frequencies more granularly than aggregate triangles, modelling the reporting delay at a finer level and conditioning on claim or policy features rather than assuming one delay pattern for an entire cohort.
The second is individual or micro-level claims reserving. Here the research is moving from chain-ladder to per-claim models that use individual claim features, claim type, severity indicators, the history of payments and status changes, and in some work reinforcement-learning approaches for dynamic, data-driven reserve optimisation that updates as a claim develops rather than re-running a static projection.
- Survival-based IBNR: models the reporting-delay distribution at claim level, sharpening frequency estimates for claims not yet reported.
- Individual-claim reserving: predicts the development of each open claim from its own features, capturing heterogeneity the triangle averages away.
- Dynamic optimisation: reinforcement-learning methods that adjust reserves as new claim information arrives.
The through-line is granularity: both strands replace a single cohort-level pattern with claim-level structure, which is where the long-tail commercial reserve has the most to gain.
Data and governance prerequisites
These methods are only as good as the claim-level data behind them, and that is the first practical constraint for an Indian insurer. Individual-claim reserving needs a clean, deep, per-claim dataset: dated payment and status histories, claim features captured consistently over time, and enough resolved claims to learn development patterns from. An insurer whose claims data is thin, inconsistently coded or only available in aggregate cannot run these models well regardless of the technique, so data readiness is the gating question.
Governance is the second constraint. A reserving method that feeds the financial statements and the regulatory capital calculation cannot be a black box the reserving committee does not understand. Under IRDAI's expectations, the appointed actuary and the reserving process carry responsibility for the numbers, so an ML reserving model has to be explainable enough to sign off, validated against held-out experience, and governed with documented assumptions and a clear account of where the model is trusted and where judgement overrides it.
Reserves, capital and the Ind AS 117 transition
Reserving in India is not standing still on the accounting side, which is why the method choice is timely. The General Insurance Council and the Institute of Actuaries of India are running IFRS 17 and Ind AS 117 implementation training for non-life insurers, which signals an active transition that affects how reserves are measured and presented.
The interaction with ML reserving runs through volatility and capital. Ind AS 117 changes the measurement and disclosure of insurance liabilities, and reserve volatility, how much the estimate moves as experience develops, becomes more visible under the new presentation. A reserving method that produces sharper, better-grounded long-tail estimates can reduce unexplained movement, but a poorly governed model that overfits can do the opposite and introduce instability, so the transition raises the premium on getting the method and its governance right.
The capital link is being built into the research directly. Recent ML and reinforcement-learning reserving work embeds solvency-capital and capital-floor constraints into the reserve optimisation, so the reserve is aligned with regulatory capital rather than estimated in isolation. That is a meaningful direction for an Indian insurer, because reserves and required capital are not separate problems; a method that optimises the reserve while respecting a capital floor keeps the two consistent.
A measured path for an Indian reserving function
The honest framing for a reserving committee is that machine learning is a complement to chain-ladder on long-tail commercial lines, not a replacement to switch to wholesale. The sensible path is incremental: identify the lines where the triangle is weakest, the long-tail commercial covers driven by a few large, slow claims, and pilot a claim-level method there alongside the existing approach, comparing the two before relying on either.
The prerequisites order the work. First, get the claim-level data into a state that supports individual-claim modelling, because without it the method cannot run. Second, build the governance, validation, explainability, documented assumptions and clear ownership by the appointed actuary, so the model can be signed off and defended. Third, run ML and chain-ladder in parallel on the chosen lines, treating divergence between them as a question to investigate rather than evidence either is right. Fourth, watch the capital and Ind AS 117 interaction, since the value of a sharper reserve is partly in steadier, better-explained movement under the new presentation.
Reserving sits downstream of underwriting and the wordings that define what is covered, so a reserving view on a long-tail line is stronger when it is grounded in how the underlying cover actually responds. Sarvada gives commercial insurance brokers and their carrier counterparts structured, searchable access to insurer policy wordings and the intelligence around them, so the claim behaviour a reserving model tries to capture can be read against the cover that generates it. Request Access to connect the reserve to the wording behind it.