Climate scenario analysis

10.5 asked what a book loses over the next three quarters if the economy turns severely adverse. Climate scenario analysis asks almost the same question over 25 years instead of nine months — three transition pathways instead of two stress scenarios, decades of checkpoints instead of quarters. It needs no new calculator to answer it.

The whole story

Did you know?

ClimateScenarioShowcase's own doc comment states the one mechanical change plainly: "unlike the stress-test page, remaining maturity is held fixed at one year at every checkpoint: this book is a steadily-rolled short-dated position being re-marked under each pathway's cumulative drift, not one trade running down to expiry over 25 years." Everything else — the ScenarioRunner.run calls, the call/put book, the seed — is identical to 10.5's. docs-reg/climate-scenario-analysis.md calls for exactly this: "if [stress testing] exists, climate scenario analysis reuses it unchanged, with the NGFS macro variables as scenario inputs" — confirmed directly against source, no new calculator anywhere.

Build the pathway grid

Three pathways, each a target spot/vol/rate shock reached via its own power curve, crossed with six checkpoints:

private record Pathway(String name, double spotTarget, double volTarget, double rateTarget, double power) {
  Scenario at(int years) {
    double f = Math.pow((double) years / HORIZON_YEARS, power);
    return Scenario.of(name + "@" + (2025 + years),
        Shock.relative("spot", spotTarget * f),
        Shock.additive("vol", volTarget * f),
        Shock.additive("rate", rateTarget * f));
  }
}

List<Pathway> pathways = List.of(
    new Pathway("orderly", -0.15, 0.03, 0.010, 1.0),
    new Pathway("disorderly", -0.30, 0.15, 0.020, 2.0),
    new Pathway("hot-house-world", -0.35, 0.10, -0.015, 1.5));

power shapes the pathway's story: orderly ramps linearly (power=1.0, smooth and early, as NGFS's own name for it implies); disorderly is quadratic (power=2.0, barely moving at first, then abrupt); hot-house-world sits between the two (power=1.5, a steady, continuous physical drag). No maturity shock anywhere in this list — the same three Shock types 10.1 introduced, just riding a different curve.

Revalue across 18 nodes

Real numbers, 500,000 scenarios per node, cpu-jit, seed 42, loss against the fixed, unshocked 2025 mark:

pathway \ year	2030	2035	2040	2045	2050
orderly	-111,919	-220,870	-326,986	-430,413	-531,344
disorderly	-42,375	-166,811	-367,007	-636,591	-971,397
hot-house-world	-127,684	-351,258	-624,889	-931,516	-1,263,832

disorderly's own numbers make its power=2.0 curve visible without needing the formula: barely -42k by 2030, only -367k by 2040, then nearly -1m by 2050 — back-loaded exactly as the name and the exponent both promise.

Did you know?

The worst projected loss here is -1,263,832, at hot-house-world@2050 — the furthest horizon, the opposite of 10.5's nearest-horizon-worst finding. Same replay mechanics, opposite shape, for the reason the first sidenote named: with maturity held fixed, there's no shrinking optionality to cap the loss as the horizon grows — only the pathway's own cumulative market drift, which never stops accumulating. 10.5's counterintuitive result and this page's intuitive one are both correct; they're answering different questions with the same tool.

Try it yourself

If 10.5's shrinking-maturity story explains why that page's losses shrink with horizon, does adding a maturity shock here make this page's losses shrink too? Built the hot-house-world@2050 node by hand, once as shipped (maturity fixed) and once with an extra Shock.additive("maturity", -0.99) layered on top, both compared against the same 2025 baseline. Real numbers: fixed-maturity loss -1,263,832; with the maturity shock added, -1,325,210 — worse, not better.

The reason is worth sitting with: 10.5's decay effect only appeared because its "loss" compared each stressed scenario against a horizon-matched baseline that carried the identical maturity shock — both sides decayed together, so the gap between them narrowed. This page's baseline is pinned to the unmoving 2025 mark throughout, so adding decay to only the stressed side just pushes it further from a reference point that never moves. Same mechanics, same ScenarioRunner.run call — but "loss vs. baseline" measures something different depending on what the baseline itself is allowed to do.

▶️ Run it

10.5's exact harness, unchanged, live — right here (the interactive example), fed the NGFS pathway grid instead of the macro/horizon one:

Or run the real thing:

mvn -o -q install
mvn -o -q -pl nablatensor-examples exec:java \
  -Dexec.mainClass=com.nablatensor.examples.ClimateScenarioShowcase

Defaults to cpu-jit, 500,000 scenarios per node, no GPU.

⚠️ What this doesn't do

The pathway curves are explicitly illustrative — the source's own doc comment says so directly — chosen to show the qualitative NGFS shape (orderly smooth and early, disorderly abrupt and back-loaded, hot-house-world a continuous physical drag), not the published NGFS macro-variable paths. The real regulatory grid steps annually, not every five years; this is an 18-node illustration of a grid a real exercise runs at roughly five times the resolution. And per docs-reg/climate- scenario-analysis.md's own "partial fit" verdict — softening as a US driver, EU/UK-only, still-maturing methodology — climate-economic scenario construction, sector transition modelling, physical-hazard mapping, and counterparty-level emissions data all sit outside a pricing engine entirely, same list the showcase's own final printout gives.

What's next

→ Deeper: docs-reg/climate-scenario-analysis.md covers the real programme detail (ECB Fit-for-55, Bank of England NGFS work, the Fed's 2025 step back) this page's code doesn't touch. → Chapter 10 (Regulatory capital) is complete. Next: CVA from first principles, a deeper look at the exposure simulation 10.3 already used for SA-CVA's sensitivity vector.