# Econometrics Econometrics is the discipline that fuses [[Economics|economic]] theory, [[Statistics|statistics]], and [[Probability_theory|probability theory]] to measure relationships in an [[Economic_system|economic system]] from data the analyst did not get to design. Ragnar Frisch coined the name in the 1920s and co-founded the Econometric Society (1930); Frisch and Jan Tinbergen shared the first economics Nobel memorial prize (1969) for building the field. Its defining problem is not curve-fitting but *identification*: an economy is a [[Feedback|feedback]]-riddled [[Complex_system|complex system]] in which prices, quantities, and policies all cause one another, so the mapping from observed correlations back to structural mechanism must be earned, assumption by explicit assumption. That makes econometrics the [[Decision_theory|decision-theoretic]] instrument panel of economics — the place where a [[Mathematical_model|mathematical model]] meets [[Time_series|time-series]] data and is forced to say something falsifiable under stated [[Uncertainty|uncertainty]]. ## Haavelmo's move: the economy as a probability draw Early skeptics argued that non-experimental data could never support inference. Trygve Haavelmo's *The Probability Approach in Econometrics* (1944) answered by treating observed data as one realization from a joint [[Probability_distribution|probability distribution]] generated by a system of simultaneous equations. That single move imported the whole apparatus of [[Estimation_theory|estimation theory]] — likelihoods, sampling distributions, hypothesis tests — into economics, and it recast every structural claim as a statement about that distribution. It also exposed the simultaneity trap: in a supply-and-demand pair, price appears on both sides, so ordinary least squares applied to either equation alone is biased. The Cowles Commission program of the 1940s–50s built the identification conditions (order and rank) that say when a structural parameter can be recovered at all — a solvability theory for inference, cousin to [[Observability|observability]] in [[Control_theory|control theory]]. ## The workhorse and its warranty The core tool is linear regression: y = Xβ + ε, with β̂ = (XᵀX)⁻¹Xᵀy chosen to minimize squared residuals — geometrically, an orthogonal projection of y onto the column space of X, pure [[Linear_algebra|linear algebra]]. The Gauss–Markov theorem states the warranty: when errors have zero conditional [[Expected_value|expected value]] and constant variance, least squares is the best linear unbiased estimator. Every clause is load-bearing, and econometrics is largely the engineering of what to do when a clause fails: heteroskedasticity-robust standard errors, generalized least squares, instrumental variables when a regressor is correlated with the error. The last is the field's signature: find a variable that shifts the regressor but touches the outcome through no other channel, and simultaneity yields — the same exclusion logic that [[Path_analysis_(statistics)|path analysis]] ([[Sewall_Wright|Sewall Wright]], 1920s) and the modern [[Causal_model|causal model]] make graphical, with causes drawn as a [[Directed_acyclic_graph|directed acyclic graph]]. ## Time series: data with memory Macroeconomic data arrive ordered, and order changes the mathematics. A [[Time_series|time series]] carries [[Autocorrelation|autocorrelation]], so effective sample sizes shrink and naive standard errors lie. Worse, trending series fool regression outright: two independent random-walk processes regressed on each other produce "significant" coefficients — the spurious-regression pathology (Granger and Newbold, 1974). The repair kit is now standard: difference or detrend to reach stationarity; model dynamics explicitly with autoregressive structures in [[Discrete_time_and_continuous_time|discrete time]] (Box–Jenkins, 1970); test for unit roots; and, when series wander together, exploit cointegration (Engle and Granger, 1987) to recover the long-run relation. State-space methods estimate latent components with the [[Kalman_filter|Kalman filter]], making macro-econometrics a sibling of [[System_identification|system identification]] and [[Signal_processing|signal processing]]; volatility clustering in finance gets ARCH models (Engle, 1982). [[Least-squares_spectral_analysis|Spectral]] and [[Frequency_domain|frequency-domain]] views of business cycles round out the toolkit. ## Policy, expectations, and the Lucas critique Econometric models are built to steer, which invites the [[Control_theory|control-theory]] temptation: estimate the system, then optimize the policy. Robert Lucas's 1976 critique showed why the analogy breaks — the "plant" is made of people. Estimated relations are equilibrium outcomes of agents' [[Decision-making|decision-making]] under a given policy regime; change the regime and the coefficients move, because expectations move. Agents are players in a [[Game_theory|game]], not components with fixed [[Transfer_function|transfer functions]], even if their rationality is [[Bounded_rationality|bounded]] in the sense of [[Herbert_A._Simon|Herbert Simon]] rather than perfectly [[Rational_choice_model|rational-choice]]. The response split the field: structural modelers estimate deep parameters meant to survive regime change, while Christopher Sims's vector autoregressions (1980) minimized theory and let the data speak. The tension — mechanism versus forecast — is the same one [[System_dynamics|system dynamics]] and [[Systems_theory|systems theory]] navigate when modeling any [[Social_system|social system]]. ## Simulation, experiments, and the modern kit Where closed-form sampling theory gives out, the [[Monte_Carlo_method|Monte Carlo method]] takes over: simulate the estimator thousands of times to learn its finite-sample behavior, bootstrap standard errors by resampling, and stress-test models the way [[Reliability_engineering|reliability engineering]] stress-tests components. The "credibility revolution" pushed design back into observational data — natural experiments, difference-in-differences, regression discontinuity — culminating in the 2021 Nobel to Card, Angrist, and Imbens. On the prediction frontier, [[Machine_learning|machine learning]] supplies regularized, high-dimensional fits, and econometrics supplies what pure prediction lacks: valid inference after selection. The division of labor is clean and useful — [[Analytics|analytics]] finds patterns; econometrics prices the [[Uncertainty|uncertainty]] and defends the causal claim feeding [[Operations_research|operations-research]] and policy [[Decision_theory|decisions]]. ## What the vault should take from it For readers arriving from systems articles, econometrics is the worked example of inference inside a loop you cannot open. It shares estimators with [[Estimation_theory|estimation theory]], state-space machinery with [[Control_theory|control]], network thinking with [[Causal_model|causal models]], sampling logic with [[Statistics|statistics]], and simulation habits with [[Monte_Carlo_method|Monte Carlo]] practice — but it adds the discipline of asking, before any fit, *what assumption lets this number mean what I claim it means*. That question transfers to every [[Complex_system|complex system]] the vault models. **On the spine:** [[Economics]] · [[Statistics]] · [[Time_series]] · [[Causal_model]] · [[Monte_Carlo_method]]. ## Wikipedia : Wikitube **Strict pair:** [Wikipedia](https://en.wikipedia.org/wiki/Econometrics) : [Wikitube](https://en.wikitube.io/wiki/Econometrics) ## Previous hub tags Hubs: `Systems`. Portals: [[PORTAL_Systems]], [[PORTAL_Monte_Carlo_method]], [[PORTAL_Decision_theory]], [[PORTAL_Reliability_engineering]], [[PORTAL_Operations_research]]. --- *Repopulated 2026-08-12 · redlink fill · 0 deletions.*