Earned Parsimony

A useful starting question is this:

That question sounds strange only if we assume economic reality lives only inside official statistical releases. The Billion Prices Project challenged that assumption by collecting large numbers of online prices and constructing high-frequency price indexes. Cavallo and Rigobon showed that online price indexes often co-move with consumer price indexes while also opening new ways to study price stickiness, international price dispersion, and real-time price behavior.

The point is not that web data make statistical agencies obsolete. They do not. Official measurement has sampling discipline, continuity, revision procedures, and institutional legitimacy. Those things matter. The more interesting point is that economically meaningful signals can live outside the canonical data pipe. A high-dimensional system may notice a pattern before the standard dashboard has a place for it.

These episodes did not prove that simple models are useless. They proved something subtler: the relevant simplicity changes when the regime changes. A variable that is ignorable in one period can become load-bearing in another.

The case for high-dimensional economics is not simply that we now have more data. The deeper case is that the economy itself is high-dimensional.

Production-network research shows why. In a world of input-output linkages, microeconomic shocks do not necessarily average away. A shock to one supplier, sector, port, commodity, technology, or region can propagate through customers, customers of customers, inventories, substitution patterns, and prices. In networked systems, the importance of a node depends not only on its size, but on its position.

That means the economy is not well-described by a universal short list of variables. The relevant variables are conditional on the target, the horizon, the regime, and the pathway of propagation. The same electricity-price movement may mean something different during a heat wave, a manufacturing boom, a war-related gas shock, or an AI data-center construction cycle.

This is why large information sets have long been valuable in macroeconomic forecasting. Stock and Watson showed that many predictors can be summarized through estimated factors when the data have factor structure. Large Bayesian VARs use shrinkage to handle richer dynamic systems. FRED-MD and FRED-QD institutionalize this idea by providing large, regularly updated macroeconomic datasets for big-data empirical analysis. The lesson is not that every variable matters. The lesson is that we should not decide too early which variables cannot matter.

The most useful analogy from language modeling is not that economics is language. It is that meaning can be learned from context.

Word embeddings popularized a simple representational insight. A word can be converted into a vector by training a model on a large corpus to predict words from nearby words, or nearby words from a word. If two words repeatedly appear in similar contexts, their vectors move closer together. The model is not handed a dictionary of meaning. It learns relationships through repeated contextual use.

The crossover to economics is natural, but imperfect. Economic variables also acquire meaning from context. A rise in diesel prices means one thing near a port disruption, another near a war-related energy shock, and another near a regional construction boom. A spike in job-search queries means one thing in a normal expansion and another during a credit contraction. A data-center permit means one thing in a region with spare grid capacity and another in a region with constrained transmission. In language, context is often a sequence of neighboring tokens. In economics, context can be built from time, geography, sector, network position, policy regime, event windows, and market structure.

This does not make embeddings causal. It makes them useful maps. A good embedding system can say: these things repeatedly become relevant together. That is not the final answer. It is a better starting point than waiting for a human to guess the connection in advance.

Economics is harder than language in several important ways. Economic data are temporal, revised, policy-sensitive, strategic, often sparse at the macro level, and deeply entangled with causality. A language model can learn that two words are related without asking whether one caused the other. A policy economist cannot avoid that question. What works for language may be useful for economics at the representation and discovery layer. It does not automatically solve causal identification, welfare analysis, or policy design.

An economic embedding system would not begin by deciding that only GDP, inflation, unemployment, and interest rates matter. It would build a broad economic corpus and let the model learn which signals carry information for a given target and regime.

The corpus would combine structured and unstructured signals. The purpose is not to worship data volume. The purpose is to reduce blind spots created by siloed datasets.

• Macroeconomic time series
• Firm-level data
• Sector data
• Regional labor markets
• Online prices
• Shipping data
• Credit conditions
• Building permits
• Energy loads
• Earnings calls
• Central-bank speeches
• News
• Search behavior
• Policy documents
• Supply-chain graphs
• Geospatial variables

A token is not only a word. In an economic corpus, a token can be a variable at a time and place, a firm in a supply network, a sector in an input-output table, a sentence in a central-bank speech, an event in a policy calendar, a price change in a product category, or a supplier-customer edge weighted by revenue exposure.

Different encoders can learn different parts of the economy. Time-series encoders can model macro and market data. Text encoders can model documents, speeches, earnings calls, and news. Graph neural networks can model supplier-customer networks or financial networks. Geospatial encoders can model regional exposure. Event encoders can model policy shifts, geopolitical shocks, or natural disasters. Training objectives could include masked economic modeling, next-period forecasting, link prediction in economic networks, contrastive learning between similar and dissimilar shocks, and event-conditioned response prediction. The model learns embeddings by being forced to predict what is missing, what comes next, or what propagates.

Attention is useful because relevance is conditional. In a transformer, attention scores how much each token should matter for a particular prediction. In economics, an attention-like mechanism could score which variables, lags, regions, sectors, documents, or network neighbors matter for a specific target.

The model does not have to believe that all variables matter. It only has to be allowed to consider them before it learns which ones to ignore.

Can computers control scope and parsimony? Yes, but only if we distinguish exploration from commitment. It is not unreasonable to let a model develop a very wide candidate space. That is exactly where computers are strong. They can scan thousands of variables, align mixed-frequency data, test many lag structures, map co-movements, track text, detect regime changes, and search for interactions that no human economist would write down first.

It becomes unreasonable only when we treat everything the model finds as truth. The right architecture is: wide search, sparse commitment. Let the computer expand the hypothesis space. Then let economic discipline, statistical discipline, and causal reasoning constrain it.

This gives economics a different kind of parsimony. Instead of human beings preselecting a small variable set because it is easy to communicate, the model is allowed to search broadly and then forced to compress under statistical discipline.

A high-dimensional economic system is not valuable because it is fashionable. It is valuable because it can perform tasks that are difficult for humans to do manually, repeatedly, and consistently.

A high-dimensional system may forecast well without explaining causally. It may predict inflation using energy loads, news sentiment, wage data, shipping delays, and policy text. That does not mean every input caused inflation. It means the inputs helped predict the target.

This distinction is central. Machine learning is often strongest at prediction. Many economic questions are about structural parameters, treatment effects, and welfare. A model can help discover candidate mechanisms, estimate nuisance functions, and improve forecasts. But policy interpretation requires identification. For causal economics, high-dimensional models should be paired with natural experiments, instruments, difference-in-differences designs, structural assumptions, double/debiased machine learning, or other frameworks that explicitly separate the target causal parameter from high-dimensional background prediction.

The strongest version of this argument is not abstract. It suggests a concrete workflow for researchers, institutions, and policy teams.

Start wide. Build the candidate corpus before narrowing. Include official data, market data, text, networks, geography, and events.

Prevent leakage. Use real-time vintages, temporal masks, and rolling-origin splits so the model cannot learn from the future.

Learn representations. Train embeddings and encoders to predict missing values, next-period movements, sector spillovers, or event-conditioned shifts.

Gate and shrink. Use attention-like relevance, sparse routing, shrinkage, and learned masks to move from broad search to conditional parsimony.

Validate brutally. Use out-of-sample tests, holdout sectors, holdout regions, placebo periods, ablations, and stress tests.

Translate into hypotheses. Ask what the model discovered. Which signals became adjacent? Which lags mattered? Which network paths carried the shock? Which variables mattered only under one regime and not another?

Identify causally. Only then move to causal designs, structural models, or policy simulations.

The AI economy is an ideal test case because its most important near-term macro effects may not sit where a parsimonious model expects them. The usual question is: how much productivity will AI create? That question matters. But a high-dimensional system would ask a second question: where is the physical infrastructure of AI already becoming an economic constraint?

AI is not only software. It is also data centers, grid interconnections, power-purchase agreements, transformers, gas turbines, batteries, water permits, construction labor, land use, local electricity prices, municipal bonds, and industrial real estate. That makes the energy question timely.

The International Energy Agency projects data-center electricity consumption roughly doubling from 485 TWh in 2025 to 950 TWh in 2030, with AI-focused data centers growing faster than overall data-center electricity use. The U.S. Energy Information Administration has also identified large computing centers as a major driver of the strongest four-year growth in U.S. electricity demand since 2000.

A narrow model might put AI into a productivity variable or a capital-expenditure variable. A wide model would ask whether the AI boom is simultaneously an energy shock, a construction shock, a regional housing shock, a utility-finance shock, and a local-price shock. It could track local data-center permits, interconnection queues, power-purchase agreements, gas-turbine lead times, transformer orders, utility capital spending, regional power prices, grid congestion, water permits, construction employment, land values, municipal bond issuance, local rent inflation, and industrial displacement. Then it could ask whether those variables begin moving together before the standard macro dashboard registers the shock.

That is the point of wide data. It does not merely add more columns to a spreadsheet. It gives the model a chance to discover that the live economic question may not be the one the modeler would have asked first.

There is a weak version and a strong version of the wide-data argument. The weak version says: more data is always better. That is false. More data can mean more noise, more leakage, more spurious correlation, more overfitting, more privacy risk, and more false confidence.

The strong version says: when the economy is a networked, heterogeneous, regime-dependent system, premature restriction of the candidate space can hide the mechanisms that matter. Therefore, the model should be allowed to explore widely under strict regularization, validation, pruning, and causal discipline. This is the position worth defending.

Let the system develop a large map of possible relationships. Then prune it. Stress it. Hold it out. Break it. Ask whether it survives future data, alternative samples, placebo periods, and causal tests. The goal is not maximal complexity. The goal is faithful compression.

Economics will always need simplification. No one can reason from an infinite model. The real question is whether simplification is imposed by habit or learned through disciplined search. The next generation of economic modeling should combine the strengths of economics and machine learning.

• Theory to define meaningful questions.
• Wide data to expand the hypothesis space.
• Embeddings to learn economic context.
• Attention-like mechanisms to gate relevance.
• Pruning to restore parsimony.
• Validation to punish overfitting.
• Causal inference to separate prediction from policy truth.