Forecasting for food price inflation using machine learning methodology: Expansion on FRED-MD

Food price inflation has become a significant policy issue in the United States again, driven by recent shocks like the COVID-19 pandemic, global supply disruptions, and rising energy and input costs. Because low-income households spend a larger part of their budgets on food, increases in food prices directly affect food security, nutrition programs, and economic policies. This study assesses various machine learning techniques for forecasting monthly U.S. food price inflation, using an extended version of the Federal Reserve’s FRED-MD dataset. We enhance FRED-MD with food- and agriculture-related variables, including fertilizer prices, energy costs, commodity market indicators, and supply chain stress measures, to better understand the drivers of food price changes. Our results show that machine learning models and forecast combination methods generally outperform simple univariate benchmarks, especially over medium-term horizons, while basic autoregressive models perform well in the short term. Combining forecasts provides the most reliable improvements, especially for six- and twelve-month inflation predictions. Additionally, incorporating sector-specific data improves RMSE across nearly all methods, although statistically significant gains are concentrated in random forest, neural networks, and forecast combination strategies, particularly for accumulated forecasts These findings emphasize the value of integrating domain-specific information into advanced forecasting tools to better support food security policy, nutrition programs, and economic planning.