Where the research stands
DSI is a research programme, and each component is stated at its own level of maturity. Overall, the protocol is specified and synthetically rehearsed: ready for shadow use, not yet evidenced for embedded reliance.
| State | Meaning | Current position |
|---|---|---|
| Demonstrated | Implemented and tested in synthetic, known-truth conditions | The frozen support rule (Wind Tunnel 1); the end-to-end case flow (IN-SIM-01) |
| Specified | Requirements and records set out in operational detail | The six steps and their records; proportionate depth; the case kit |
| Mapped | Questions and interfaces identified | Combining uncertainty across claims impact steps; calibrating quantitative bias adjustments; institutional placement |
| Exploratory | Early hypotheses | Adaptation to public purchasing; extensions beyond the current protocol |
Building on actuarial evaluation of care management
Between 2004 and 2006 the Society of Actuaries Health Section published a series of papers on evaluating care management interventions. It set out how actuaries should measure the financial outcomes of such programmes and named the threats that matter most: selection into the programme, regression to the mean, underlying trend, drop-out and claims runout.
The DSI protocol builds on that measurement work and addresses the step that follows: how far a measured effect may enter a specific pricing, reserving or reinsurance decision, within which limits, and what reopens it.
How large is the effect, and is the comparison sound?
How far may it change this decision, for how long, and what reopens it?
A synthetic test of the support rule
Wind Tunnel 1 tested a frozen support rule in synthetic worlds where the true effect was withheld from the procedure. Tolerances for false recognition and wrongful denial were declared in advance, 144 candidate rules were narrowed to one, and the frozen rule, a Poisson difference-in-differences design with employer-clustered uncertainty, was evaluated on 7,000 held-out simulations.
The last result matters most: a conservative bound cannot cure a biased comparison. Identification comes before estimation.
The whole protocol, end to end
A synthetic insurer of 60,000 members across 40 employers, with engagement-driven selection, incomplete linkage and immature claims, was taken through all six steps and repeated across 200 simulated worlds.
The bound covers chance, not selection
The frozen design’s lower bound overstated the true effect in all 200 runs, while its own pre-trend diagnostic flagged the cause each time.
Fallback and residual bias assessment are both needed
The agreed fallback removed most of the bias; it still exceeded the truth in 56% of runs.
Timing is first-order
Ignoring the lag before effects emerge doubled the claims impact.
Monitoring needs its reporting lags
A fall in persistence reached the decision owner three months late.
Use limits held
The record stayed within its declared decision.
These findings shaped Protocol Manual v0.2.1.
Research in progress
Identification on observational data
Stress-testing comparator construction, selection, drift and fallback designs, with the residual bias assessment declared in advance.
Robust estimation
Inference with few or dominant clusters, dependence and robust estimators under the support rule.
Claims impact
Combining uncertainty across the steps from supported effect to expected claims, including effect emergence and transferability.
Monitoring triggers
When later evidence should maintain, increase, reduce or withdraw a recognition.
Next: an adversarial stop-case under the frozen v0.2.1 rules, independent expert review, and a shadow assessment on a real programme.
Presentations
| Date | Forum | Contribution |
|---|---|---|
| 8–13 November 2026 Upcoming | 33rd International Congress of Actuaries (ICA 2026) ↗ Tokyo International Forum, Japan · IAA and Institute of Actuaries of Japan | Paper, AFIR-ERM programme: Avoided Claims, Evidence Risk and the Wind Tunnel: a prevention-evidence recognition protocol for insurance economics |
| 19 August 2026 | Asian Actuarial Conference 2026 (AAC 2026) ↗ Marina Bay Sands, Singapore · Singapore Actuarial Society | Health track: Prevention: can it really bend the cost curve? How would you know the curve bent? |
| 18–22 May 2025 | 2nd Joint Colloquium of IAA Sections (JoCo 2025), Actuaries in Sustainability ↗ São Paulo, Brazil · IAA Sections, hosted by the Instituto Brasileiro de Atuária | Presentation, published by the IAA: Embedding Prevention in Sustainable Insurance Solutions ↗ |
Researchers, actuarial bodies, insurers, reinsurers and health economists interested in the evidence-to-decision gap are welcome to write. Papers are shared on request while the programme develops.