27 Machine-Learning Models Tried to Forecast US Inflation — a Classical One Won

Why forecasting CPI-U matters

CPI-U is one of the most consequential numbers in the U.S. economy. It's built from roughly 80,000 monthly price quotes collected across around 23,000 retail and service businesses plus 50,000 rental housing units, weighted to reflect what urban consumers actually buy: food, energy, housing, medical care, transport, and more. It covers an estimated 93% of the U.S. population. The Federal Reserve leans on CPI readings when it sets the federal funds rate, and the same number feeds into wage negotiations and cost-of-living adjustments across the economy. A model that forecasts it well has real, practical value beyond academic interest.

Twenty-seven models, one dataset

The researchers — Jassimar Singh and B. S. Kambo — used a Python forecasting toolkit (PyCaret) to fit 27 different time-series models to monthly CPI-U data going back over two decades, sourced directly from the U.S. Bureau of Labor Statistics. The lineup ranged from simple baseline forecasters, to classical statistical methods like ARIMA and exponential smoothing, to more modern machine-learning approaches including random forests, gradient boosting variants, and regularized regression methods. Before fitting any model, they tested whether the data behaved the way these models assume: they confirmed the CPI-U series is not random noise (it has real, statistically significant patterns over time), is not normally distributed, and is not stationary in its raw form — meaning its statistical properties shift over time rather than staying constant. To fix that last problem, they applied a standard transformation (differencing) that made the series suitable for modeling.

The result: a classical method beat the machine learning field

Every model was scored on three standard error metrics — how far off its predictions were, on average, from what actually happened. By the most commonly used of these (root mean square error, or RMSE), the ETS model came out on top, narrowly ahead of plain exponential smoothing, and clearly ahead of ARIMA. It also beat every machine-learning model in the comparison by a wide margin, including random forests and several gradient-boosting variants.

What the forecast actually says

Using the winning ETS model, the study projected CPI-U forward two years, from January 2024 through December 2025. The annual average moves from about 304.7 in 2023 to roughly 314.1 in 2024 (a rise of about 3.1%) and about 323.6 in 2025 (a further rise of about 3.0%) — continued inflation, but at a fairly steady, moderate pace rather than accelerating or collapsing. Within each year, the forecast shows a recurring seasonal shape: prices climb more quickly from January through July, then flatten out or fluctuate only slightly for the rest of the year. The study offers several plausible, non-exclusive explanations for that pattern — tax refunds boosting spring spending, seasonal demand for travel and warm-weather goods, agricultural prices moving with planting and harvest cycles, and general economic activity picking up after the winter months — without claiming to isolate which factor dominates.

Why the simplest model might be winning here

There's a broader lesson in why a decades-old statistical method outperformed newer machine-learning techniques on this particular series. CPI-U has a strong, fairly regular trend and seasonal pattern and comparatively little noise — exactly the conditions classical exponential-smoothing methods were built to handle well. Machine-learning and deep-learning models tend to earn their keep on messier, more irregular, higher-frequency data (stock prices, for instance) where flexible, non-linear pattern-matching pays off. A single, well-behaved macro time series with a clear trend and seasonality is arguably closer to the kind of data ETS was designed for than the kind of data that rewards a more flexible model — which may be part of why the classical approach won here, even though it wasn't tested against every other series it might face.