The model put to the test
How we check that the game holds up, and how its responses compare with those of published studies.
In plain words
We first run every country for a hundred years, with and without shocks, to check that no figure becomes absurd and that the accounts add up every day. We then repeat the experiments of published studies, such as a rate rise or a fiscal stimulus, in order to compare the game's responses with theirs. Finally, we play France, one decision at a time, then with six strategies over eight years.
The checks, over 213 countries and 100 years
Two worlds are run. One is without shocks or elections, where a fault cannot hide behind bad luck. The other is with its shocks, elections and defaults, the world of a game.
| Check | Without shocks | With shocks |
|---|---|---|
| no NaN or infinite value | passes | passes |
| all variables within their bounds | passes, except the out-of-domain countries (below) | only breaches due to bad luck, at most 0.007% of country-months (tolerance 0.1%) |
| the debt moves exactly by its flows, minus haircuts | passes (10⁻¹⁴) | passes (10⁻¹⁴) |
| every euro paid is received: the rows and columns of the matrix sum to zero | passes (10⁻¹⁶ of the day's largest payment) | passes |
| the debt owed is the debt held | passes | passes |
| economies settled over the last ten years | passes: 98% of countries settled, 3 still converging and 1 in a cycle of defaults (below), none drifting away | — |
| two runs with the same seed are identical to the bit; another seed gives another world | — | passes |
| a country run absurdly (revenue 10%, spending 80%, rate 0): no NaN, and no other country moves by a single bit, shocks, elections and events included | — | passes (no link between countries; trade will create one) |
| the list of out-of-domain countries is up to date: each one still fails somewhere | passes (2 countries) |
Countries still converging. After a hundred years without shocks, three countries have not finished returning to their anchors, but they are getting closer: Lebanon and Gaza, whose debt is still coming down, and South Sudan, whose inflation is at 5.5% for a target of 5.3%. The check reports them separately. It fails for a variable that drifts away, or for a country that has never defaulted and does not converge. Before monetary transmission was corrected (E1), there were eleven of them, Argentina among them. Sudan, which starts from 263% debt, defaults again and again, even without shocks. A confidence crisis drawn at random pushes it into default as long as its debt stays in the risk zone. The check reports it separately.
Countries out of domain (§6): Macao, and Zimbabwe, which starts from 687% inflation (December on December 2024), beyond the hyperinflation bound. The model has neither money nor monetary financing, and these are what make and unmake hyperinflations. Sudan has left the list.
Against the literature
| Experiment | Model | Literature |
|---|---|---|
| spending +1 point of GDP, output gap by year (advanced) | 0.70; 0.89; 0.53; 0.26; 0.12 | IMF template scaled to the same peak: 0.71; 0.89; 0.54; 0.27; 0 (Batini et al. 2014) |
| tax cut of one point, first year | 0.23 (advanced); 0.28 (emerging) | 0.25; 0.2 to 0.4 (Batini et al. 2014) |
| policy rate +1 point for one year, output | −0.26% in the 18th month | advanced: −0.25% (corrected for publication bias) to −1% (raw), around 18 to 24 months (Enzinger et al. 2025); emerging: weaker than in advanced countries (Enzinger et al. 2025), or −1.05% in the 10th month (Brandão-Marques et al. 2020) |
| same, price level after 5 years | −0.065% (advanced); −0.13 to −0.15% (emerging) | −0.15% (corrected) to −0.9% (Enzinger et al. 2025; Havranek and Rusnak 2013) |
| cost of a one-point disinflation | 1.43 (France), 1.45 (United States), 1.01 (Brazil), 0.96 (India) | 1.4 on average on quarterly data, 0.8 on annual data, from 0 to 3.6 (Ball 1994); 1.3, and 1.7 once corrected for publication bias (Enzinger et al. 2025) |
| effective rate − nominal growth, at the start (medians) | advanced −2.6; emerging −3.6; low-income −6.8 | 1999-2008: G7 +1.3; emerging −4.2; developing −9.2 (Escolano et al. 2017) |
| volatility of growth and inflation (E13) | simulated / observed: 1.06 and 1.02 at the median; from 0.93 to 1.23 for growth, from the 10th to the 90th percentile | the small shocks alone: 1.00 and 0.99, by construction; world events (E18) come on top of the ordinary times that the target measures |
| years of negative growth (annual averages) | advanced 22%; emerging 13%; low-income 7% | 2000-2019: 18%; 12%; 12% (World Bank) |
| recessions dated quarter by quarter (advanced) | 3.5 per decade, 30% of the time, 3 quarters, −3.2% | 1.2 per decade, 11% of the time, 3 quarters, −1.9% (Claessens, Kose and Terrones 2008) |
| elections lost by the incumbent camp | 47% of the 4,954 elections in a hundred years | a government loses 2¼ points per parliamentary term on average (Nannestad and Paldam 1999); US incumbents won 8 times out of 15 after one term, 2 times out of 8 after two (Abramowitz 2008) |
| sovereign defaults | 36 in a hundred years, in 15 countries, i.e. 0.17% per country per year | 180 restructurings from 1970 to 2010 in 68 countries (Cruces and Trebesch 2013), several times more |
| world events (E18), cost in the first year | severe pandemic −6.35% (France), −6.53% (Brazil); solar storm −4.86%; El Niño −0.99% (Indonesia); earthquake −1.10% (Vanuatu); cyclone −1.08% (Philippines); drought −1.39% (Botswana) | their sources: −6.3; −6.6; −4.75; −1.01; −1.11; −1.11; −1.41 |
| end of a high inflation | without shocks, Argentina, starting from 118%, is at 56% after ten years and below 25% after nineteen; Turkey goes from 44% to 23% in ten years and 18% in twenty, for a target of 10% | crises above 40%: 6 years at the median (Bruno and Easterly 1998); episodes above 100%: 40 months on average (Fischer, Sahay and Végh 2002); inflations of 15 to 30%: 31 episodes out of 55 last three years, only 6 more than five (Dornbusch and Fischer 1993) |
One decision at a time, on France
France is played, without shocks or elections. Each decision is compared with the same France in which the player touches nothing. Differences from this baseline:
| After | Revenue −2 points: output gap | purchasing power | popularity | debt | Spending +2 points: output gap | unemployment | popularity | debt | Rate +2 points: output gap | inflation | debt |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 6 months | +0.52 | +3.18 | +1.20 | +0.2 | +1.57 | −0.22 | +0.80 | −1.2 | −0.11 | 0.00 | +0.2 |
| 1 year | +0.68 | +3.29 | +1.62 | +0.8 | +2.03 | −0.50 | +1.36 | −1.3 | −0.36 | −0.01 | +0.7 |
| 2 years | +0.49 | −0.13 | +0.11 | +2.8 | +1.46 | −0.53 | +0.21 | +0.2 | −0.83 | −0.05 | +2.4 |
| 5 years | +0.06 | −0.05 | −0.02 | +10.1 | +0.17 | −0.07 | −0.06 | +7.2 | −1.26 | −0.13 | +10.4 |
At ten years, the differences no longer measure the decision, because France left to itself defaults in February 2034, at 156% of GDP. Its primary deficit of 4%, kept up, takes it into the risk zone, where a confidence crisis drawn at random brings it down. Non-player countries escape this, since their fiscal rule closes that deficit.
Readable? Yes. The three decisions go in the right direction, with orders of magnitude that the literature recognises, and their cost comes after their effect. The tax cut makes the government popular, for one year: +3.3% of purchasing power, +1.6 points of popularity, then nothing more, because purchasing power is measured over twelve months. So does a spending stimulus, now: +1.4 points after one year, through the benefits it pays (E14) and through jobs (−0.5 points of unemployment). The rate rise, when it is held, weighs more and more, because its effect comes slowly (E1).
Who wins and who loses in play
Six strategies played with France, each over forty seeds, for eight years (2026-2033). The player holds the presidency, and loses when their party no longer holds either the presidency or the government. What the orders do, and what the player goes through in power. This table is old. Measured again, it has changed a great deal. The National Assembly rejects a good share of the orders, and no strategy leads to default any more.
| Strategy | Loses in 2027 | Loses in 2032 | Survives eight years | Goes through a cohabitation | Goes through a default | Final debt, median |
|---|---|---|---|---|---|---|
| do nothing | 14 | 23 | 3 | 8 | 0 | 125% |
| rate at 0 from the start | 13 | 24 | 3 | 8 | 0 | 116% |
| austerity from the start (revenue +2, spending −2) | 16 | 22 | 2 | 7 | 0 | 115% |
| stimulus sixteen months before the vote (revenue −2, spending +1.5) | 12 | 26 | 2 | 10 | 9 | 123% |
| the same, eight months before | 10 | 26 | 4 | 11 | 6 | 138% |
| big stimulus eight months before (revenue −5, spending +4) | 3 | 33 | 4 | 27 | 37 | 125% |
No game is lost through a default. In France, the fall of the government leaves the president in place, and opens a cohabitation.
What it shows. If it does nothing, France loses 2027 one time in three. The recessions that the world brings (E13, E18) do not spare it. Timing matters. The same stimulus loses 2027 less often when it is placed eight months before the vote rather than sixteen, because purchasing power is measured over twelve months. Two strategies give two stories. Played cautiously, you lose at the ballot box, in 2027 or in 2032, and you go into cohabitation only one time in five. Played as a big spender, you win 2027 almost for sure, then the debt leads to default (37 games out of 40), cohabitation follows two times in three, and you lose 2032. Almost everyone loses in 2032. After two terms, Abramowitz's time-for-change penalty (−4.3 points) is rarely made up. The big stimulus no longer dominates. It survives eight years four times, as often as the others, after an almost certain default. Before cohabitation, the same table let nobody survive. Re-elected in 2027, the player lost the 2029 legislative elections 24 times out of 26, under the time-for-change penalty of a presidential election, and this defeat ended the game. Over forty seeds, each cell is uncertain by three or four games. Smaller gaps do not tell the strategies apart.