The model

The model runs four economies, the United States, the euro area (as France and the rest of the union, on one central bank rate), Sweden and China, month by month, starting from each as measured to the end of Q3 2026. Each run goes on to the end of June 2032, longer if the bubble bursts later, so that every run covers the same length of time after the bust. Where the data measure what the model does, the charts show them before the start.

Choose how governments respond and how large the bust is, how fast it falls and on which day share prices peak. The page runs the model on your choice.

How it works

Each month, spending in each economy responds to what happened the month before:

The AI investment bust. Share prices fall, investment stops and lenders lose capital, on the dot-com crash's own measured path. US AI investment is measured net of imported computers (0.4 points of GDP below its 2015–22 trend), so in the US the bust has no investment leg: a collapse in it falls on the countries that make the equipment.
Lost wages. People who lose their jobs spend less, which costs more jobs. Unemployment benefits offset part of this; an income guarantee replaces most of the lost wage.
Jobs lost for good. In a downturn, firms automate some of the jobs they cut instead of rehiring. AI also takes jobs steadily, whatever the economy is doing. Office work can be automated from the start; physical work only once robots arrive. The central pace comes from the research; on their trend since Sep 2024 the US jobs data allow at most 22% of it (39% measured from Sep 2024 to the latest three months of data).
Leaving the workforce. In the United States the model assumes most people whose jobs AI takes stop looking for work: 94% of them, the share of the fall in employment since Sep 2024 that took this form. That share rests on one month of data; measured from other starting points it is 18–59%. It changes how the loss splits between unemployment and leaving the workforce; the number out of work is the same either way. They lose the same income, so the model counts them in spending, bank losses and benefits; the unemployment rate does not count them. Jobs lost in the downturn itself are counted as unemployment. Outside the United States every job lost counts as unemployment.
Banks. Loan losses reduce bank capital, and banks with less capital charge more for credit. Rescues limit the losses. Without deposit insurance, bank runs shrink the money people spend.
Trade. A slump in one economy cuts the exports of the others. Tariffs cut them further.
Prices and debts. Falling prices raise the real burden of debts, up to a limit fitted on past downturns. Central banks cut rates, but not below their floor.
Policy on dates. Each government response starts on a date: the month unemployment had risen by a set amount in the central case, the same whatever the bust.
Care. Health care and social assistance keep hiring at their usual pace unless budgets are cut. With funding, the state can hire more, up to Norway's share of jobs in care.

One set of parameters was fitted by month on nine past downturns and checked on the dot-com bust, which it was not fitted to:

Limits

Many settings are approximations, each stated in the model's code. It shows what can happen and why. It is not a forecast.
It has no exchange rates, so it overstates Sweden's exposure.
It recovers from 2009 faster than the United States did, so it is optimistic about how long a slump lasts.
How far falling prices deepen a slump is only loosely pinned down by the downturns it is fitted on.
The share out of work cannot rise above 50% and output cannot fall more than 60% below capacity in the model; a result at either limit means the slump in the model had not reached its lowest point.
The bust can be up to twice the size of the dot-com crash, when US household stock wealth falls 89% at the bottom, about as far as share prices fell in 1929–32, and it can fall from half to twice as fast. The model was fitted on nine past downturns, from 1929–33 to 2011–13, and checked on the dot-com crash; a run outside what those tested says so.

Model as of 1b3d593, 4 Nov 2026. The code and every result are public at github.com/wilsoniumite/labor (in paths/code/global_crash.py).