Rethink required: Why economists became enamoured of mathematics—and what, if anything, the numbers are good for

For more than a century, economics has been inching closer to mathematics—first cautiously, then all at once. From post‑war general equilibrium proofs to the modern machinery of dynamic stochastic general equilibrium (DSGE) models, empirical causal inference, and machine‑learning‑aided forecasting, the discipline often speaks in equations before it speaks in prose. To many outsiders that looks like a fixation; to many insiders it looks like progress. As 2025 pushes economics into an era shaped by data firehoses and generative AI, the question is not whether math belongs, but why economists fell so hard for it—and which parts still earn their keep.

The romance began with a promise: coherence. Early formalists argued that if economics aspired to be a science, it needed axioms, theorems, and results that did not wobble with rhetoric. Proofs supplied that backbone. Models translated fuzzy stories about incentives into tractable systems that could be solved, tested, and compared. In classroom after classroom the catechism spread: specify preferences, constraints, and technology; write down an objective; derive optimality; study equilibria. The mathematics conferred discipline in two senses—an academic field found its method, and students found their posture.

But the method also narrowed what counted as knowledge. If a question could not be modeled cleanly—bounded rationality in messy institutions, norms that shift, power that bites—it often moved to the footnotes. Stylised agents in frictionless markets were mathematically convenient even when empirically rare. Many papers optimised respect from peers instead of welfare in the world, proving existence and uniqueness while dodging measurement, implementation, or politics. As one graduate joke had it: if reality refuses to fit, so much the worse for reality.

A countercurrent emerged with data. The credibility revolution of the 1990s and 2000s reframed identification as the central task of empirical work. Difference‑in‑differences, regression discontinuities, instrumental variables, and later synthetic controls promised estimates of causal effects that were legible beyond the seminar room. Randomised controlled trials in development economics—backed by increasingly careful ethics reviews—put fieldwork and design on equal footing with algebra. Here, mathematics served as audit trail: assumptions were explicit, estimators transparent, standard errors properly interrogated.

Then came computation. Econometrics picked up new tools and old caution. Machine learning offered predictive power; economists insisted on interpretability and counterfactuals. The uneasy compromise—‘use ML to predict, then structure to explain’—now defines practice from tax compliance to labour‑market matching. Central banks feed high‑frequency indicators and firm‑level microdata into hybrid models that splice DSGE skeletons with data‑driven muscles. The COVID shock taught humility and agility; inflation’s 2021‑2023 rollercoaster taught that expectation formation and supply constraints cannot be tacked on as afterthoughts.

If mathematics is a language, the critique is not that economists speak it, but that they sometimes mistake fluency for wisdom. A beautifully identified local average treatment effect may be exquisitely true and trivially useful if it lives too far from policy margins. Conversely, messy structural models can guide real‑time decisions—so long as their misspecifications are known, communicated, and monitored. Usefulness hangs on institutional detail, error bars, and plumbing: what lever can a minister actually pull; how fast does a central bank’s balance‑sheet change transmit; who pays when an algorithm errs?

Consider the wins that mathematics delivered. Auction theory, nurtured in the greenhouse of stylised game models, re‑designed spectrum sales and saved taxpayers billions. Market design matched doctors to hospitals and kidneys to patients with fewer failures. Mechanism design clarified why some regulations backfire. Pricing carbon—still politically fraught—rests on a quantitative spine: damage functions, discount rates, and general‑equilibrium spillovers. Industrial‑organisation models armed enforcers to challenge mergers that would have quietly dulled competition in digital markets. Even basic budgeting now leans on probabilistic forecasts instead of point guesses. These are not footnotes; they are infrastructures.

The misses matter too. Over‑confident DSGE models declared the financial system stable until it wasn’t; calibration often disguised ignorance as precision. Some quasi‑experimental studies amplified a narrow sliver of reality into sweeping claims—credible identification became a shield for incredible extrapolation. P‑hacking and publication bias haunted literatures until pre‑analysis plans and registries pulled them back. Forecasting competitions embarrassed bespoke macro models; nowcasts chased noise. None of this invalidates mathematics. It indicts a style of doing economics that confuses optimisation on paper with implementation in practice.

Where does generative AI fit? In 2025, large language models can code estimators, simulate policy games, and summarise literatures in minutes. They also hallucinate and inherit biases from the texts they ingest. The frontier is not replacing economists but re‑allocating their scarcest resource—attention. Math helps here by forcing clarity about objects (treatments, outcomes), estimands (ATE, quantile effects), and welfare criteria. When an AI proposes a fix to a bottleneck in tax collection, a formal model can separate clever from credible, and a pre‑registered evaluation can separate credible from causal.

Policymaking has also become more ‘ops’. Central banks now communicate in fan charts and scenario trees. Finance ministries run distributional impact calculators before unveiling budgets. Competition authorities simulate platform fee changes with agent‑based models. City halls use origin‑destination matrices to redesign bus routes overnight. These practices depend on mathematics but thrive on humility: a loop of deploy‑measure‑update, with dashboards that display uncertainty rather than bury it in appendices.

So what should change? First, elevate the craft of measurement. Too many models take inputs as given when measurement is the model. Price indices that mis-handle housing, productivity metrics that ignore intangible capital, or carbon ledgers that miss upstream emissions will tilt any optimisation. Mathematics is not just optimisation; it is also errors‑in‑variables, missing data, and unit definitions. When measurement improves, policy relevance follows.

Second, tilt incentives toward external validity. Reward studies that map where an estimate travels and where it breaks. Design papers should ship with ‘operational readmes’: what data feeds the tool, what governance keeps it honest, what failure modes trigger a rollback. Mathematical elegance is a feature; robustness is a deliverable.

Third, pluralise methods without relativising standards. Agent‑based models, systems dynamics, and qualitative institutional analysis often speak past mainstream econometrics. They should speak with it. A pluralist economics still needs shared benchmarks: pre‑specified outcomes, reproducible code, transparent priors, and loss functions that reflect public values, not just squared errors.

Finally, reform the social contract of graduate training. The canonical sequence—real analysis, optimisation, econometrics—builds muscles but not always judgement. Curricula that embed field exposure, policy labs, and interdisciplinary rotations create economists who can talk to engineers, lawyers, and sociologists without translating every sentence into Greek letters. Mathematics remains the spine—but vertebrae are not the whole body.

So yes, economists are enamoured of mathematics. The better question, in 2025, is whether they are enamoured of the right maths for the right jobs. When theory clarifies trade‑offs, when evidence identifies causal levers, and when models are wired into the noisy circuits of institutions, mathematics is not an affectation. It is how a complex society reasons under constraint. The rethink required is not to demote equations, but to insist they earn their place at the table—tested, translated, and tied to consequences.