Macro Climate Regimes and U.S. Sector Rotation

Abstract

Practitioners have long believed that the economic "climate" determines which equity sectors lead the market, yet the link is usually drawn through a single growth-and-inflation grid such as the well-known investment clock. This paper asks whether a richer, machine-learned description of the macroeconomic environment can do better. The answer, for the most part, is that it cannot, although the analysis identifies the one application where it genuinely earns its keep. We apply k-means clustering, a simple and transparent unsupervised-learning algorithm, to eleven monthly U.S. macroeconomic indicators spanning growth, inflation, monetary policy, the yield curve, credit spreads, financial conditions, and household sentiment, sorting every month from January 1986 through June 2026 into five recurring "economic climates." All inputs are public, drawn from the Federal Reserve Economic Data (FRED) service, the Kenneth French data library, and exchange-traded fund (ETF) prices, and every result is reproducible from the accompanying replication package. The five climates align closely with the credit cycle and with familiar episodes: the 2008 and 2020 crises fall in a credit-stress contraction, the 2021–2023 inflation surge in late-cycle overheating, and the early-1990s and 2001 downturns in recessionary easing. Each climate carries a distinct sector-rotation signature: technology leads in mid-cycle expansions, energy and defensives lead in overheating, and rate-sensitive defensives lead in easing. The dispersion between the best and worst sectors also widens dramatically in the adverse climates. Two disciplined tests, however, temper the machine-learning promise. First, in a complexity-matched comparison, the algorithmic climates do not describe forward sector returns any better than the investment clock: a five-state clock built from only growth and inflation matches the eleven-variable clustering in sample, and out of sample the simple clock is more accurate. Second, a strictly real-time strategy that tilts toward each climate's historically leading sectors fails to beat an equal-weight sector portfolio net of transaction costs. The durable application is risk control rather than sector selection: holding the equal-weight sector portfolio in the three benign climates and moving to cash in the two adverse ones roughly halves the maximum drawdown over 2001–2026, to 23 percent against roughly 50 percent for the buy-and-hold alternatives, while matching the market's return and improving the Sharpe ratio from 0.55 to 0.64. This capital-preservation benefit survives an untouched 2011–2026 holdout, a grid of alternative exposure settings, and, critically, reconstruction of the entire macro dataset from real-time ALFRED vintages exactly as the data appeared on each historical decision date. It is competitive with, though not clearly superior to, the best simple alternatives such as trend following, and unlike them it comes with an economic diagnosis an investor can inspect, debate, and override. As of June 30, 2026, the framework reads the U.S. economy as a mid-cycle expansion, a benign climate in which the strategy is fully invested. The broader lesson is methodological: honest out-of-sample design, complexity-matched benchmarks, and real-time data vintages are what separate a genuine empirical finding from an artifact of hindsight.