We introduce a new methodology for identifying high-frequency fiscal shocks using Large Language Models. We apply this method to 1947–2025 US data. Our results show that the model successfully mimics a “professional forecaster” of the current and future US fiscal position, and is able to recover similar shocks to what have already been identified in the narrative fiscal shock literature. We then examine the effects of fiscal shocks on asset prices: in response to a 1pp shock to the present-value of the current and next ten-years deficits, ten-year Treasury yields rise more than 30bps, with real yields and break-even inflation expectations both contributing to the rise. The dollar appreciates significantly — as much as 4.8% — and the term-premium, as measured by the 2Y-10Y spread, rises 16–24bps. Turning to macroeconomic outcomes, our fiscal shocks produce government spending multipliers in the 0.5–1 range. Tax shocks show strong signs of anticipation, and using our data to account for anticipation, we find that output and consumption fall by more than 2% in anticipation of a 1% of GDP tax cut. The multiplier for an anticipated tax shock is −1.2, smaller than typical estimates.

We use survey data on macroeconomic expectations, across 89 countries and going back to 1989, to establish four facts about how forecast biases depend on the time horizon of the forecast. The data cover average expectations and horizons from 0 to 10 years. (1) Expectations underreact at a horizon of one year or less. (2) Expectations overreact at horizons of two years or more. (3) Expectations are “too extreme” at all horizons. (4) Overreaction and over-extremity increase with forecast horizon. These four patterns hold across advanced and emerging economies, and across multiple macroeconomic variables. They are inconsistent with several popular models of overreaction, where the degree of overreaction is independent of forecast horizon. However, we show that a model featuring costly recall, uncertainty about the long-run mean, and sticky-information can match all four of our facts. Finally, although long-term expectations exhibit stronger overreaction, it is short-term expectations that are most strongly associated with fluctuations in GDP, investment, and the stock market.

We study how financial market prices can be used to forecast the likelihood of transformative artificial intelligence. Transformative AI is a double-edged sword: while advanced AI could lead to rapid economic growth, some researchers argue that superintelligence misaligned with human values could pose an existential risk to humanity. Theoretically, we show that either possibility would predict a large increase in long-term real interest rates, due to consumption smoothing. We then use rich cross-country data on real rates and growth expectations to show that, contrary to other recent findings, higher long-term growth expectations are indeed associated with higher long-term real interest rates. We conclude that monitoring real interest rates is a promising framework for forecasting AI timelines.

Over the last 30 years, the correlation across emerging market countries' sovereign debt spreads is more than double the correlation in their GDP (0.67 vs. 0.33). This discrepancy suggests that movement in sovereign spreads is primarily driven by global factors, not local fundamentals. Using data for 38 emerging market countries, I confirm that global variables are far more significant — have more than an order of magnitude larger R² — than local variables in explaining spread movement. Further, as evidence of the importance of price of risk channels for explaining spread movement, the share of a country's debt that is held by foreign investors significantly predicts the sensitivity of the spread to global financial conditions. I then build a three-period multi-country sovereign default model. The model shows that “standard” model features alone only produce spread correlations between 0.3–0.4. Introducing either cyclical investor risk-aversion or cross-country connections in variable costs of default matches the empirical correlation of 0.67. Yet, spreads in the model remain more tightly linked to fundamentals than in the data.

We decompose the post-1973 productivity growth slowdown into three causes: structural change (Baumol's cost disease), input misallocation, and pure productivity effects. We do this by constructing sector-level productivity from 1947 to 2016, using the recent BEA-BLS Integrated Level Production Accounts adjusted with our own industry-specific markup estimates. We find that Baumol's cost disease explains ~25% of the productivity slowdown. The magnitude of the input misallocation channel is sensitive to methodology, reflecting uncertainty about whether or not aggregate markups have risen: input misallocation can account for between 0–20% of the productivity slowdown. Finally, we also show that linear growth fits the US data better than exponential growth, though under either exponential or linear growth there has been a post-1973 productivity slowdown.