Fertilizer to food.
A causal chain from natural gas to consumer CPI, with the empirical lag every link actually shows.
Food inflation reports are too lagged and too aggregated to forecast much. The variables that actually drive grocery prices a year out sit further upstream, in markets that don't usually get grouped together: gas, ammonia, fertilizer, grain futures. We wanted to know whether the chain shows up clearly enough in the data to be predictive, or whether it's mostly hand-waving. So we built it.
3–6 wk · 1–3 mo · 2–4 mo · 4–6 mo end-to-end
Twenty-five years of monthly IMF commodity prices and US BLS CPI, 2000–2025, about 300 observations. Run monthly log-returns through cross-correlation and Granger tests across every plausible lag. The headline number: IMF Fertilizer Index leads US CPI Food at Home by five months at correlation 0.30, Granger p of 0.004, significant at every lag from one to twelve.
Where does the five-month lag come from? Mostly two places. Ammonia plants run about a month of feedstock inventory, so urea spot reacts to gas with a three-to-six-week delay. And retail food is sticky — supplier contracts insulate processors, processors hedge their inputs, and even after wholesale moves, repricing a shelf has menu costs of its own. The middle of the chain is faster than people expect: grain futures are forward-looking, so they price the planting response within weeks of a fertilizer shock instead of waiting for the actual harvest.
One thing surprised us. Most of the relationship lives in the tails. 2008 and 2021–2022 drive most of the correlation; in quiet years the signal disappears into noise. So this isn't useful as a continuous regression model — but it's useful as an alerting layer. Watch the upstream nodes when they move hard. Ignore the chain when they don't.
Potash → soybeans is the weakest pair in the table. Correlation 0.12, below significance for our sample. That's not the analysis failing; it's the biology. Potassium isn't the marginal nutrient for soy the way nitrogen is for corn, so you wouldn't expect a clean pass-through. A model that predicts which links work and which don't is more useful than one that finds spurious strength everywhere.
The chain isn't a trade. It's a list of nodes worth watching when the upstream moves.