Tools (models)¶
The model bench. Choose the actuarial models that turn your soft data into hard results — by hand, or let Scelo suggest a set for your data.
The Tools workstation: the dataset hub, attached model nodes, and the model catalog. The animated illustration needs JavaScript.
The canvas¶
Tools is a typed node graph that reads the way an Unreal Engine Blueprint does: inputs on a node's left, outputs on its right, and a pin only where something actually flows.
- The Dataset Hub is the source. It has one output pin for each role your
data can really play — claims triangle, mortality table, model points,
claim counts, claim amounts, rating factors, target & features,
exposure, numeric columns, WMTR parameters — with its evidence beside
it (
7 origins × 7 dev periods, or the columns it reads). Each role is found by the same check the model runs in Hard, so a pin on the hub is a promise: if the hub offers a claims triangle, chain ladder will find one. Once models are attached, the hub shows the roles they read and folds the rest under ▸ N more this data can feed. Click a role for the models that read it. - Model nodes carry their family and name in a tinted header, then their pins, then a one-line rationale. A model that can't run here says so under its rationale — ⚠ can't run: needs a claims triangle — with the full reason on hover.
- The catalog in the right-hand panel lists every model by family, marked ✓ runs on this data, ~ illustrative, or ✗ its inputs aren't here. Click a model to attach it; click it again to detach it.
The banner above the canvas is the check before you move on — for example ✓ 4 ready · 1 off · 2 wires, or ⚠ 3 ready · 1 can't run. Switched-on models whose inputs are all fed run in Hard; the rest report why not.
Pins and wires¶
| Pin | Means |
|---|---|
| Round | Data from the Dataset Hub |
| Diamond | A result from another model |
| Hollow | Nothing plugged in yet |
| Dashed, red | A required input nothing can feed — the model can't run |
Pins and wires take the colour of what they carry. An optional input shows the default it falls back to beside its label — built-in defaults, severity only, no rate stress. A required input that's missing says not in this data or nothing plugged in; when another model could feed it, the pin offers the fix instead: + GBM (LightGBM), switch on … or plug in ….
Every input that takes a role your data has is fed from the hub automatically (unless a model is wired into it), and those wires can't be unplugged. Model-to-model wires carry results, and only four exist — each one consumed by the model it feeds:
| From | Into | What flows |
|---|---|---|
| Lee-Carter or Cairns-Blake-Dowd | Life Contingencies | projected mortality, priced as a cohort |
| GLM · frequency | GLM · severity | claim frequency — × severity gives the pure premium |
| GBM (LightGBM) | SHAP explainability | fitted GBM — SHAP explains exactly this model |
| Economic Scenario Generator | SCR · Standard Formula | rate scenarios, sized into an interest-rate stress |
The reserving methods take no model inputs: Mack and the bootstrap refit chain ladder themselves, and Bornhuetter-Ferguson keeps a prior of its own, because a chain-ladder prior would collapse it onto chain ladder.
New models arrive wired, and so do models you switch on. A model that joins the canvas — from the catalog, the picker or the chat — or is switched on is plugged into what it can consume or feed, taking the first switched-on model that makes it (Lee-Carter before CBD); wires you drew or unplugged are left alone. A fresh pick (identify models, regenerate) wires every model it picks. A model that can't run without another brings it along: adding SHAP adds the GBM it explains.
Connecting¶
Drag from an output (right) to an input (left). A wire joins only pins whose types agree, and an input takes one wire — so plugging CBD into Life Contingencies unplugs Lee-Carter. While you drag, a tag at the cursor says what the wire carries and, over a pin, whether the drop will take (✓ release to plug in) or why not: ✗ claim frequency can't feed mortality — it takes a mortality table or projected mortality, or ✗ a model can't feed itself.
Drop a wire on empty canvas and a menu opens with the models that fit: the ones that take what the wire carries or, dragged from an input that takes a model's result, the ones that make it. Picking one adds it and plugs it in. Right-click the canvas for every model, ranked by what this data can feed. Both menus group their entries under fits this data, illustrative, on the canvas and can't run on this data (greyed out, with the reason); type to filter, ↑/↓ to move, Enter to pick, Esc to close.
Dragging a hub pin onto an input that a model currently feeds hands it back to the dataset (✓ release to feed it from the dataset instead). To unplug a model-to-model wire, click the × on it, or select it and press Backspace or Delete. A wire whose source is switched off turns dashed, and its input falls back to the dataset or its default, if it has one.
Choosing models¶
By hand — click a model in the catalog, or add one from the canvas menus (see Connecting):
- Reserving — Chain Ladder, Mack Chain Ladder, Bornhuetter-Ferguson, Bootstrap (IBNR).
- Mortality / longevity — Lee-Carter, Cairns-Blake-Dowd, Life Contingencies.
- Pricing / GLM — GLM · frequency, GLM · severity, GBM (LightGBM), SHAP explainability.
- Forecast / capital / climate — WMTR forecast, Economic Scenario Generator, CLIMADA climate hazard exposure.
A model whose figures come from built-in assumptions rather than your data carries an illustrative tag on its node: SCR · Standard Formula, Economic Scenario Generator, DB / DC Valuation, Smith-Wilson · curve fit and Economic curves. A model the data can't feed can still be attached, but it arrives switched off.
AI-suggested — click identify models. Scelo reads your dataset's shape and domain and proposes a set, with a short rationale per pick. You can accept, add to, or swap them. Picks the data can't feed arrive switched off, with the reason, so Hard won't run them unless you switch them on; a pick that needs another model brings it along (SHAP arrives with its GBM).
Per-model controls¶
Each model node has:
- A scoped chat (
ASK SCELO ▸) —swap chain-ladder for Mack,explain this model's assumptions,compare models. It wires, too —wire cbd into life contingencies,unplug lee-carter from life contingencies— as long as both models are already on the canvas. - ↻ to replace it with another model — the ones that read the same inputs are listed first, each marked ✓ fits or ✗ can't run — an on/off switch, and × to remove it (or select it and press Backspace / Delete).
- Click a node for its details in the right-hand panel: whether it can run, each pin and what feeds it, and why it was picked.
A model's theory, run output and diagnostics are in its detail dashboard on the Hard stage.
Model notation renders mathematically — e.g. a WMTR rationale reads "αM / αT / αR triplet detected".
Other actions¶
| Action | What it does |
|---|---|
| identify models | AI-suggest a model set for the data |
| regenerate | Re-run the AI suggestion |
| re-layout | Lay the graph back out left to right and refit the view |
| export · code | Export the model setup as a script |
| ← back: soft / next: hard → | Move through the pipeline |
The canvas lays itself out left to right until you drag a node; from then on it keeps your arrangement, and a model you add by dropping a wire or right-clicking lands where you did it. re-layout hands the layout back to Scelo.
When your model set looks right: next: hard →.