Project Record

Reinsurance Model Run Observability

Operational dashboard and lineage layer for multi-model reinsurance runs on Tyche—tracking outputs, runtime, treaty KPIs, and parameter sensitivity.

Domains

ReinsuranceInsurance

Capability

Workflow Infrastructure

Methods

WorkflowObservabilityData IngestionMonte CarloVisualization

Context

As actuarial analyst on a reinsurance desk, I supported five to six treaty, reserving, and exposure models that batch-ran on Tyche, a leading actuarial software platform. Stakeholders needed a single view of what each run produced, how long it took, and whether outputs were consistent across models—not a rewrite of the pricing engines, but operational intelligence on top of them.

Problem

Model outputs landed as opaque directory trees with no unified lineage. Run duration, CPU use, and cross-model agreement were invisible until someone manually backtracked files. When treaty parameters changed—retention, quota share, attachment, or limit—there was no fast way to compare before/after performance or spot which model diverged.

Architecture

Ingest: PowerShell scanners on Windows compute nodes walk Tyche output paths; parallel fan-out by model folder and treaty year.
Parse: Python normalizes run metadata, file manifests (row counts, checksums, header schemas), and resource telemetry into an indexed store.
Visualize: Refreshable dashboards show runtime percentiles, CPU/memory use, file-dependency graphs, and actuarial KPI overlays.

Reinsurance Metrics Overlay

Treaty KPIs are derived from model outputs, not re-implemented pricing logic:

Ceded loss ratio: $\text{CLR} = \sum L_{\text{ceded}} / \sum P_{\text{ceded}}$ — compare across runs and models.
Net retention: $R = 1 - P_{\text{ceded}} / P_{\text{gross}}$ — sanity check on treaty structure.
XL utilization: $\min(1,\; L_{\text{layer}} / L_{\text{limit}})$ — attachment and limit stress.
Geographic ceded exposure: choropleth by region and line of business layered on the ops view.

Parameter Sensitivity and Cross-Model Comparison

One-at-a-time sensitivity on treaty parameters $\theta \in \{\text{retention}, \text{limit}, \text{quota share}, \text{attachment}\}$ :

\Delta_k = \frac{M(\theta_k + \delta) - M(\theta_k)}{\delta}

where $M$ is ceded premium, net loss ratio, or runtime. Tornado charts rank drivers; Monte Carlo sweeps perturb $\theta$ within actuarially plausible ranges and report mean and p5/p95 of net loss ratio and ceded premium—using cached outputs or scaling factors where models are linear in quota share, and full reruns where they are not.

Cross-model agreement uses normalized distance on shared dimensions (LOB, region, treaty year):

d_{ij} = \frac{|M_i - M_j|}{\max(\epsilon,\; \tfrac{1}{2}(|M_i|+|M_j|))}

Pairs above tolerance surface for review, analogous to reconciliation breaks on actuarial metrics rather than positions.

Speed and Refresh Techniques

Incremental indexing: watermark on mtime and path; scan only new or changed outputs since last refresh.
Content-hash deduplication: skip re-parsing identical artifacts across reruns.
Dependency-aware refresh: run DAG; if upstream manifest is unchanged, skip downstream KPI recompute.
Metadata-first parsing: lightweight manifests before heavy actuarial file reads.
Pre-aggregated rollups: nightly snapshots of per-run KPIs so dashboards query summaries, not raw outputs.
Selective tile refresh: ops metrics, geo map, and sensitivity panels refresh independently.

Full-tree scans dropped from tens of minutes to sub-minute incremental refresh on typical daily batches.

Trade-offs

KPI aggregation is heuristic and model-version dependent. The sensitivity layer supports what-if exploration, not full stochastic reserving. Tyche path conventions require maintenance when model versions or folder layouts change. The dashboards accelerate diagnosis and treaty oversight but do not replace actuarial sign-off on model methodology.