Austan Goolsbee on Central Banking as a Data Dog
Chicago Fed president Austan Goolsbee talks with Tyler Cowen about how a working central banker actually reasons: why he calls himself a ‘data dog’ rather than a hawk or dove, why he refuses to pre-commit to a monetary rule, whether post-pandemic inflation was demand or supply, the run risk in stablecoins and shadow banking, the decades-long puzzle of housing and construction productivity, the case for a decentralised Fed, and why the training-sample problem makes him sceptical of AI magic.
Key ideas
- Be a data dog, not a bird. Goolsbee rejects the hawk/dove framing for an empiricist’s creed — ‘there’s a time for walking and a time for sniffing’. The most useful thing academic economics gave him as a Fed president is the discipline of getting into the data and thinking hard about causality and identification.
- Never make policy off an accounting identity. He will not pre-commit to a monetary or nominal-GDP rule: MV = PY has ‘no economic content’, and a rule premised on past demand shocks fails when supply shocks dominate — as in the 2023 ‘golden path’, when inflation fell almost as much as it ever has in a year without the recession every rule demanded.
- Post-pandemic inflation was not all demand. Three puzzles resist the demand-only story: inflation soared while unemployment was above 6 per cent; it rose simultaneously in domestically-driven economies with far less stimulus; and it fell only when supply chains healed, not when stimulus rolled off.
- Shadow banking is the central financial-stability conundrum. Money-like deposits redeemable on demand need deposit insurance or serious asset restrictions, or they end ‘in bank runs and tears’ — ‘500 years of financial history’ say so. With official banks only about a fifth of lending, the Fed should stay the ‘fuddy-duddy’ and keep risk outside the payment system.
- The training-sample problem limits both rules and AI. An LLM — or a rule-bound central bank — is only as good as its past data, so it hallucinates confidently when the present is novel: trained on history, it would have demanded a recession in 2023. Where a mistake is ‘unforgivable’, hallucination is disqualifying.
Content
Central banking as a data dog
Asked what academic economics turned out to be useful at the Fed, Goolsbee reaches immediately for identity: not a hawk, not a dove, but a member of the ‘data dogs’ — ‘there’s a time for walking and a time for sniffing, and knowing the difference’. The discipline of getting into the data, and economics’ habits of causality and identification, are what he values. Cowen presses him to teach the connection between money and inflation, and Goolsbee lays out competing models — the FRB/US model’s Keynesian ‘inflation from overheating’, the old monetarist ‘M2 correlates with inflation’ — but says he starts from basic supply and demand and a micro sense of causality, asking first whether a move is a supply or a demand shock. The M2-to-inflation relationship broke down, he argues, because financial innovation (electronic payments, cards displacing cash and cheques) unmoored the velocity of money: with both M and V moving, over-indexing on M misleads. When Cowen tries to draw him toward a nominal-GDP rule, Goolsbee refuses — an accounting identity has ‘no economic content’, and rules only work if future shocks resemble past ones. His standing counter-example is the 2023 ‘golden path’: every forecaster and every demand-driven rule predicted recession, yet inflation fell almost as much as it ever has in a year without one.
Was post-pandemic inflation demand or supply?
Cowen argues the inflation was mostly demand — prices stayed up, services rose sharply, M2 rose 40 per cent. Goolsbee declines to say Cowen is wrong but poses three puzzles for a pure-demand story. First, US inflation soared while unemployment was above 6 per cent and output below potential — soaring inflation with slack should give a demand-only theorist pause. Second, inflation rose simultaneously across countries with far less fiscal and monetary stimulus than the US. Third, inflation did not come down when the stimulus rolled off, but did come down in 2023 as supply chains healed. When Cowen offers Japan and Switzerland as demand-side evidence, Goolsbee bats it away — those are import-heavy and had low inflation for decades before COVID — and redirects to large, domestically-driven economies like Nigeria and India that also saw high inflation at once. His settled position: not 100 per cent supply, a serious demand component, but the demand-only account is unpersuasive.
Stablecoins, shadow banking, and the lender of last resort
Cowen pushes to the theoretical edge — why does money matter at all, if it is a near-substitute for T-bills? Goolsbee concedes economists cannot fully explain it; there is some ‘yield to safety’ markets price that theory does not capture, which makes M2 an even shakier guide in a world of money-substitutes. On stablecoins fully backed by treasuries he is ‘not anti-crypto’ but nervous: money-like deposits redeemable on demand need either deposit insurance or serious restrictions on their backing, or they end ‘in bank runs and tears’ — ‘500 years of financial history’ warn of it. This is the core financial-stability conundrum: an overseen official banking sector alongside a growing, run-prone shadow sector the Fed neither controls nor has information on — and formal banks are only about one-fifth of lending. He agrees with Cowen (citing Jeremy Stein) that squeezing official banks too hard drives activity into the shadows through regulatory arbitrage, but declines the stronger claim that bank capital rules are therefore useless: after March 2023 the big banks were not the problem, which he credits to the capital raised since 2008. Returning to the theme under CBDCs later, he invokes the free-banking era’s recurrent panics (1847, 1857), Paul Volcker’s question of what sits inside versus outside the rescue safety net, and Eric Budish’s paper on the 51-per-cent attack, concluding — via Edmund Burke — that a central bank must stay ‘the fuddy-duddy’ and keep risky things off the payment system: ‘I like safe.‘
Housing, construction, and the productivity puzzle
Cowen floats a revisionist claim: maybe 2008 was not a nationwide housing bubble, just an anti-bubble panic. Goolsbee had not framed it that way, but from touring his Midwest district he reports the dominant complaint is ‘despair’ about housing cost — and not only in zoning-constrained cities: in rural Iowa the binding constraint on attracting workers was that workers could not afford housing. Housing’s relative price has risen roughly 4–5 per cent a year for a decade and a half, well before COVID, compounding into a large gap he finds genuinely hard to explain — which lends some weight to Cowen’s suggestion. His paper with Chad Syverson finds construction productivity flat-to-negative over long periods: ‘somehow, we’re forgetting’ how to build homes. Ed Glaeser’s land-use-regulation account (and a lost mid-century move toward modular, prefab building) is a candidate, but Goolsbee documented what the puzzle is not without finding a smoking gun for what it is.
The case for a decentralised Fed
Running the Chicago Fed is like running a corporation, and Goolsbee credits a professional COO — the ‘first vice president’ — for much of what looks like presidential competence; the useful micro ideas are the theory of delegation and the distinction between cost, opportunity cost, and marginal cost (things with a marginal cost of zero, like sharing information, can be done generously — knowledge as a public good). Asked whether fiscal pressure might one day cut the 12 regional research staffs, he defends the 1913 design’s ‘genius’: 7 politically-appointed governors plus 12 Reserve Bank presidents chosen by regional boards, so the FOMC — which he calls ‘the world’s greatest deliberative body’ — is not just New York and Washington. Consolidating to one data-driven staff, however efficient it sounds, invites monoculture and groupthink; better to ‘let some flowers bloom’. Cowen bargains him down from twelve toward five or three; Goolsbee holds the line with a baseball analogy — the game is worse when all the teams are in New York.
AI: productivity, hallucination, and the limits of magic
On AI’s five Fed functions — monetary policy (the ‘opposable thumb’), the payment system, bank supervision, financial services to banks, and regional expertise — Goolsbee sees a plausible role in flagging supervision risks, but notes a standing rule against general AI use over confidential-data leakage, and doubts an on-premises model would beat humans at predicting bank failures: ‘you’re only as good as the training sample’, which in 2023 would have demanded a recession. On growth, he separates one-time level effects (hybrid work, the Great Resignation, a jump in new-firm creation) from durable technology-driven growth; if AI is a general-purpose technology like electricity or computing, gains diffuse over decades from producing sectors to using ones, and Chicago Fed work finds the current surge concentrated in AI-intensive industries — consistent with an early pattern, but adoption still too low to explain much. He flags an over-anticipation risk of overheating if equity values and AI-infrastructure investment outrun realised productivity. On o1 Pro he distinguishes look-up questions (AI wins) from judgement questions with no correct answer, and calls hallucination disqualifying where a mistake is ‘an unforgivable error’. He resists the ‘just use the paid model’ rebuttal with a dot-matrix-printer analogy, and warns via self-driving cars that rapid improvement can slow sharply once gains depend on more data and compute. He divides the world into those who ‘believe in magic’ and those who do not, and places himself firmly among the sceptics. The conversation closes on a lighter note — how a young Goolsbee out-debated a national-champion Ted Cruz by teasing him until he lost his temper.
See also
- Austan Goolsbee — speaker
- Tyler Cowen — host