We are building a digital twin of the economy: households, firms, banks and governments simulated one quarter at a time, in the open, and tested against the best forecasting models we can find.
Your spending is someone else's income; your loan is someone else's asset. A digital twin keeps every one of those ledgers at once, then runs the world forward. Making that trustworthy takes a model, a test, and a way to look inside.
An agent-based model of the United States. Every firm, household and bank carries its own income statement and balance sheet; markets clear by search and matching, so shortages and rationing are things you can watch, not assume away. Forked from the Bank of Italy's BeforeIT.jl and rebuilt on public data: BEA, BLS, the Fed.
US_BeforeIT.jl on GitHub · Julia · Apache-2.0
Forecasts from 61 quarterly origins since 2010, 500 free-running paths each, scored on identical cells against nineteen other columns: naive rules, autoregressions, a Bayesian VAR and two DSGE models re-estimated at every origin. No behavioural parameter ever sees a forecast error. Every claim ships with its significance test and its caveat.
A simulation is a black box that emits indicators, until you can open it. The Economy Complexity Explorer zooms from the whole economy to a sector to one firm's ledger: who paid whom, for what, and how it added up. Five prototype lenses so far, all rendered from the same saved runs.
The engine runs fifteen phases per simulated quarter. Here they are in seven moves. Nothing is solved for an equilibrium: agents act, markets match, and the accounts settle.
Numbers below are read straight from the committed result files. They are a diagnostic on revised data, not a real-time track record. We would rather show you the caveats than hide them.
The width of the growth fan is model dispersion, not simulation noise, and in backtest its 90% band covers about what it should. The inflation band is known to be too narrow; read it as a lower bound on uncertainty.
Nine of the twenty scored columns. The repaired model tops the field in all three sample tracks and is significantly better than both DSGE columns; its edge over the simplest statistical benchmarks is marginal, and we say so. Its interval coverage on inflation still fails our own bar.
The published US input–output tables do not balance once bridged across valuation bases. Left unreconciled, the opening accounts silently drained 2.75 points from growth, mostly through one-sided rationing. Accounting first, behaviour second.
Swapping the entire 2024 industry structure for 2017 or 2012 moved the headline score by at most 0.0007. A frozen opening labour market cost a whole target. Getting the starting state right matters more than the wiring diagram.
Exploding investment booms, too-narrow inflation bands and a lost pandemic-track score all trace to how agents look ahead. Heterogeneous, survey-disciplined expectations are next on the list.
Condition the twin on the lockdown, trade, fiscal and monetary shocks that actually happened, give it the demand-pull pricing rule from the Bank of Canada's CANVAS model, and start it in December 2020. It produces the inflation surge that contemporaneous official projections put near two per cent. Give the same shocks to cost-push-only pricing and you get 1.3%.
Two-year rise in the consumer price level from the December 2020 information date. Known defects are registered: wage indexation has no real-wage brake, and the policy rule never sees the model's own inflation. A separate paper, when it is ready.
One modelled firm stands for many real ones. Slide the scale to see what each rung buys and what it costs. Getting to one-to-one is the point of the project, and most of the engineering.
Step costs measured in Julia 1.10 on an Apple M-series laptop; 99% of a step is the goods market's search and matching. A batched Rust backend for that one market is being explored.
We never claim the twin is better in general. Each rung has a preregistered protocol, the strongest challenger able to produce the same object, and an honest-failure clause: a lost rung is published with the same tables as a won one.
Prototype lenses over the same saved runs, being unified into a single explorer with deep links and semantic zoom. Exact traced transactions, model-equation flows and derived estimates are always labelled separately.
Side projects that test pieces of the vision at a scale where you can see every agent.
Every household owns one house on a small grid, carries a full balance sheet, and decides whether to move, buy or sell through a language-model decision module with hard budget and mortgage constraints. Watch prices, turnover and inequality emerge month by month.
Draw your income, expenses, assets and debts as nodes on a canvas, then ask an assistant to project them years forward. The same stock-flow thinking as the twin, for one household at a time.
The core is open source on purpose. Simulation software that won in engineering did so by making the smallest useful claim, proving it repeatedly, then expanding. We want collaborators who like that discipline.