Know the Economy, Not the Data Lake

How many indicators should a GDP nowcast use, and how should that number be chosen? A common answer is to add every available series and let a factor model or a machine-learning method separate signal from noise. We argue that this misplaces the effort for most economies. The number of indicators that can help is bounded by how many distinct sources of cyclical signal an economy generates, which is a feature of its structure rather than of the available data.

We write the choice as a bias–variance trade-off. The accuracy-maximising count rises with the sample length and with the number of orthogonal signal dimensions, and it falls with volatility, so less-diversified economies nowcast best with fewer indicators. We test this on twelve economies over a common 2000–2026 out-of-sample window.

The optimum is smaller in emerging economies (mean 3.2 against 6.4; p = 0.026), and it is explained by economic complexity (0.61) rather than income (0.35, not significant). Norway is the identifying case: rich but resource-concentrated, its optimum sits with the emerging economies, where complexity predicts it and income does not.