Huge, fancy AI models are actually worse at predicting the economy than the basic tools we were using ten years ago.
Despite being trained on billions of data points, 'parameter-heavy' neural networks fail on economic data because they lack the specific human-designed features that account for how indicators like unemployment actually move over time.
Machine Learning vs Foundation Models for Economic Forecasting Evidence from Blue Chip Indicators
SSRN · 6432986
Foundation models pre-trained on billions of time series promise universal forecasting capabilities, yet their performance on low-frequency economic data with limited observations remains untested. <div> We evaluate ten forecasting models across four families (machine learning, deep learning, extended LSTM, and foundation models) on nine U.S. economic indicators using 49 years of Blue Chip consensus forecasts (1976-2025). We employ walk-forward validation with expanding windows and Diebold-Maria