27 Machine-Learning Models Tried to Forecast US Inflation — a Classical One Won
- ▸A study comparing 27 machine learning and statistical models found a classical exponential-smoothing method (ETS), not machine learning, best forecasts US CPI-U inflation.
- ▸The winning model projects continued moderate inflation through 2025 — roughly 3% annual increases — with prices rising faster in the first half of each year and leveling off in the second half.
- ▸A companion study by the same authors found a different model (LSTM) won when the identical method was applied to India's CPI-IW, showing model choice doesn't transfer automatically across countries.
Every month, the U.S. government publishes a number that quietly shapes interest rates, wage negotiations, and cost-of-living raises: the Consumer Price Index for Urban Consumers, or CPI-U. It tracks what a typical household pays for everything from groceries to rent to healthcare, and the Federal Reserve watches it closely when deciding whether to raise or cut interest rates.
A new study asked a simple question: if you throw dozens of different forecasting techniques at this number — from simple statistical methods to more elaborate machine learning approaches — which one actually predicts it best? The researchers tested 27 different models against more than two decades of CPI-U data and measured how close each one's predictions came to what actually happened.
The winner wasn't the most complex option. A classical statistical technique called ETS (short for Error, Trend, Seasonality) — a method that has been used in forecasting for decades — edged out every machine learning and deep-learning-style model in the comparison, including popular ones like random forests and gradient boosting. Using that winning model, the study then projected where CPI-U is headed through the end of 2025: continued, moderate inflation, rising a little over 3% a year, with a recurring seasonal pattern of faster increases in the first half of each year that levels off in the second half.
One of this paper's own authors, B. S. Kambo, is a former Deputy Director (Statistics) at India's Ministry of Labour & Employment, and the very same journal issue carries his companion study applying this identical 27/28-model comparison approach to India's own Consumer Price Index for Industrial Workers (CPI-IW). There, a neural network (LSTM) won instead of ETS, projecting India's CPI-IW to rise roughly 3.4 points in the summer/monsoon months before stabilising over winter. Read together, the two studies are a useful reminder for India's own inflation-forecasting work (used by the RBI and MoSPI): the best forecasting model is series-specific, not universal, so methodology validated on US or global data cannot simply be assumed to transfer to India's price data.
Frequently Asked Questions
What is CPI-U?
CPI-U is the U.S. Bureau of Labor Statistics' Consumer Price Index for All Urban Consumers -- a monthly measure of what urban households pay for a basket of goods and services (food, housing, energy, medical care, transport, and more). It covers about 93% of the U.S. population and is the inflation gauge the Federal Reserve watches most closely when setting interest rates.
Which model best forecasts US CPI-U inflation, according to this study?
ETS (Error, Trend, Seasonality), a classical exponential-smoothing model, had the lowest forecast error (RMSE) of all 27 models tested -- narrowly ahead of plain exponential smoothing, and clearly ahead of ARIMA and every machine-learning model in the comparison, including random forests and gradient boosting.
Why did a classical statistical model outperform machine learning models here?
CPI-U has a strong, fairly regular trend and seasonal pattern with comparatively little noise -- exactly the kind of data classical exponential-smoothing methods were designed to handle well. Machine-learning models tend to have an edge on messier, more irregular data, so a clean, well-behaved macro series may simply favor the classical approach.
What does the study forecast for US inflation in 2024 and 2025?
Using the winning ETS model, the study projects CPI-U's annual average rising from about 304.7 in 2023 to roughly 314.1 in 2024 (about +3.1%) and about 323.6 in 2025 (about +3.0%) -- continued, moderate inflation rather than acceleration or a sharp slowdown.
Why does US inflation tend to rise faster in the first half of the year?
The study points to several contributing, non-exclusive seasonal factors: post-holiday spending picking back up, tax-refund-driven spending between February and April, seasonal demand for travel and warm-weather goods, and agricultural prices tied to planting and harvest cycles -- without isolating a single dominant cause.
Does the same forecasting method work for India's inflation data?
No. In a companion study by the same author team, published in the same journal issue, a Long Short-Term Memory (LSTM) neural network -- not ETS -- was the best-fitting model for India's Consumer Price Index for Industrial Workers (CPI-IW). That shows forecasting-model choice is series-specific and doesn't automatically transfer from one country's data to another's.
What data and tools did the researchers use?
Monthly CPI-U data (base 1982-84 = 100) from the U.S. Bureau of Labor Statistics, spanning more than two decades through December 2023. The 27 models were fit and compared using PyCaret's time-series module in Python, scored on Mean Absolute Error, Root Mean Square Error, and Mean Absolute Percentage Error.
Primary Sources
Cite This Article
EconoLens Research Desk. (2026, August 1). 27 Machine-Learning Models Tried to Forecast US Inflation — a Classical One Won. EconoLens. https://econolens.co.in/news/ml-models-forecast-us-inflation-cpi-ets
The EconoLens Research Desk reviews academic papers in economics and econometrics, translating cutting-edge research into accessible analysis. Full credit is given to original authors in every review.