How Prometheus stress‑tests a household

The engine takes a setup: a household's investments in equity, fixed income and property, its cash savings and salary, the rare but devastating risks specific to it (a medical bill, critical home damage, career loss) and a ruin threshold. From that it produces a distribution of year-end wealth across one million simulated paths.

At the heart of the engine is the historical record. Prometheus is trained on more than fifty years of economic history across 18 countries, calm stretches and crisis years alike, and every simulated path learns from it how markets, inflation, interest rates, house prices and jobs move together.

This matters because risks rarely move in isolation: a high inflation shock can trigger a rise in interest rates, which in turn hurts equities, which can deepen into a recession and end in a job loss. Conventional simulators do not capture these chains.

Every path is a complete year with all of those risks in play at once. Hidden risks, currency exposure, inflation shocks, fat-tailed yield-curve moves, are part of the stress from the start, not bolted on afterwards. The platform then reports the full distribution at key percentiles and gives the user the tools to interrogate it: what-if analysis, tail-risk driver decomposition, conditional probabilities, and threshold sweeps.

Markets behave in characteristic ways. Stocks have calm stretches and panic stretches. Yield curves bend, steepen and flip in a small handful of recognisable patterns. Inflation drifts, then jumps. We don't want a simulator that flattens all of that into a smooth bell curve or a linear projection, because that isn't the world money lives in. So each market risk gets its own purpose-built calibration.

Equities. For each calibrated index, S&P 500, FTSE 100, FTSE All-Share, NASDAQ, MSCI World, we fit an AR-GARCH model on decades of monthly returns. That mouthful does two things that matter: it remembers that quiet months follow quiet months and stormy ones follow stormy ones. Because it draws shocks from a fat-tailed distribution, a once-in-a-decade drawdown is never treated as impossible. How deep the bad years go is learned from the long record of 18 countries' stock markets, so an index with a short or lucky history is not taken at its word.

Currencies. Currency pairs (USD/GBP, GBP/EUR, USD/EUR) get the same treatment. If any part of a household's wealth sits in a foreign currency, the exchange rate is a risk in its own right, and we model it with the same care we apply to equities.

Bonds and yield curves. Rather than guessing one number for "the bond market", we look at the whole curve, every maturity from short bills to long bonds, and ask what its moves actually look like. The answer is that they almost always reduce to three patterns: the whole curve shifting up or down, the curve steepening or flattening, and the middle bowing relative to the ends. We calibrate each of those patterns with a heavy-tailed distribution so the simulator can produce realistically jagged rate scenarios, not just gentle parallel drifts.

Inflation. Same idea. The inflation term structure moves in a small number of recognisable shapes, and we calibrate each one. The simulated inflation rate then flows into each path's expenses, so when a high-inflation world arrives in the simulation, that year's bills actually go up, and the household feels the squeeze on the way to year-end.

Property. Residential property returns are calibrated per region (UK, US, EU) from the long-run record. Property is slower-moving than equities, but it isn't immune, and pretending it is would understate the size of a bad year. How far house prices can fall is likewise learned from many countries' housing busts, not one region's.

The market isn't the only thing that can sink a year. A job loss, a serious medical bill, a critical home repair, these don't show up on Bloomberg, but they show up on a household's balance sheet. Prometheus treats them as first-class citizens of the simulation, on equal footing with the market factors.

For each life risk described, career loss, plus whichever custom risks matter to the household, the engine asks two questions on every path: did it happen this year, and if it did, how bad was it? The "how often" piece comes from a frequency assumption (some risks fire at most once, others can recur). The "how bad" piece is calibrated from the household's own optimistic and pessimistic loss estimates, so the severity range reflects its situation rather than a textbook average.

If a risk is insured, the engine applies the cover the way the policy would in real life: a deductible per event, an annual ceiling on what the insurer will pay, and the rest hits the household's wealth. Risks sharing a policy share its limit. The point is to show what the insurance actually buys, and, just as importantly, where it runs out.

If risks were independent, finance would be easy. A bad equity year would be cushioned by a normal property year. An inflation spike would arrive on its own, not riding alongside a rate hike. Job losses would happen in random years rather than clustering with recessions.

That isn't the world we live in. The worst years are bad precisely because shocks arrive together. So rather than assume how risks relate, Prometheus learns it from the record. The engine is trained on more than fifty years of economic history across 18 countries: the oil shock of the 1970s, the recessions of the early 1980s and 1990s, the dot-com bust, 2008, the pandemic and the rate shock of 2022, with all the ordinary years in between.

That history isn't decoration. Inside the simulation, it is what makes every path a single, internally consistent world. When a simulated year turns ugly for equities, the other risks don't roll their own independent dice: house prices, jobs, inflation and interest rates move the way the engine has learned they move in years like that. A path where equities crash and inflation surges and the household's career risk fires isn't ruled out; it's exactly the kind of compound bad year the engine is designed to find, and it turns up far more often than in a model that treats each risk on its own.

Life events are tied in too. In Prometheus, a job loss is about a quarter more likely in a year when unemployment surges than in a calm year, a link the engine measured on history rather than one it assumes. For any other risk, a household can say what it tends to arrive with, a market crash, falling house prices, rising rates, another life event, and how tightly.

Press simulate, and the engine plays the household's year out a million different ways. Each of those million paths is a complete, self-consistent story: an equity return, a bond P&L, an inflation rate that drives the actual expenses, a property return, currency moves on any foreign holdings, and whichever life risks did or didn't fire. At the end of the year, each path has a wealth number, and we ask the question that matters: did the household breach its ruin threshold, or didn't it?

The summary on screen - median, p10, p5, p1 - is just the distribution of those million year-end wealth values, sorted from worst to best. But the percentiles aren't the interesting bit. The interesting bit is what's behind them.

That's why every percentile is inspectable. Open p5 and it doesn't just say "the year ended at X", it shows the specific combination of shocks that led there. Maybe equities were down 28%, inflation ran at 8%, the career risk fired, and the bond ballast didn't save the year. That's a story to act on: it shows which exposure was load-bearing in the bad years and which wasn't. It turns a number into a narrative.

And the million paths don't just sit there. They're the raw material for the real work: quantifying the tail and finding out what breaks the plan. On the platform the picture can be reshaped - pull a lever, change an allocation, insure a risk - and the distribution responds against the exact same simulated world, so the comparison is honest. The slice of paths that match a scenario worth worrying about can be carved out and the probability of ruin read under it. The bust subset can be decomposed into which factors were doing the damage, not in the abstract, but in that household's specific year. The point isn't a prettier dashboard. It's a sharper answer to the only question that matters: how much loss can a household absorb, and what's most likely to deliver it?

The horizon is one year. This is not a multi-year retirement projection or a financial-planning calculator. It does not forecast markets, recommend products, or say what to buy. What it does is something deeper, and, we'd argue, more useful: it shows how today's position behaves under the joint, fat-tailed shocks of the world as it actually is. One question matters more than the rest, how bad a household's year can really get, and what would break it. Prometheus was built to reveal both. The methodology behind it powers institutional stress-testing engines at banks, insurers and regulators, turned, for once, on a household's balance sheet.