AI nearly matches economists in live GDP forecasting contest
Czech National Bank Governor Aleš Michl said AI nearly matched economists in live GDP forecasts run since 2023. Average errors were 0.22 percentage points for Random Forest alone and 0.21 for economists using models and judgment.
A thousand models compete
At Kyoto University, Michl described a live comparison using the same data, forecast day and target quarter, rather than historical backtesting.
The human team uses all available models plus judgment; its competitor is a single Random Forest without human adjustment.
That method combines a thousand simple forecasting models using random slices of data.
Short-term forecasting began during covid with a weekly index using electricity consumption and highway truck traffic.
Machine learning followed in 2023.
AI now also helps write code and retrain models.
Sensitive data stay on internal GPU servers running open-source models; rules govern cloud use.
The bank also uses AI to forecast inflation, while humans retain responsibility for judging results.
Higher rates and larger reserves
Michl said inflation fell from 17.5 percent when he became governor in mid-2022 to 2 percent in February and March 2024.
He cited latest CPI inflation of 2.4 percent and a 4 percent policy rate.
His strategy keeps rates above inflation over the long term, restrains money growth and calls for a smaller budget deficit.
A stronger koruna lowered import costs without currency intervention, he said.
Reserves total about $180 billion, or 44 percent of GDP; the equity share doubled to 26 percent and gold increased.
The bank completed nationwide, round-the-clock instant payments in 2025.
Parity is not superiority
The live contest establishes near parity, not an AI victory, on the reported average errors.
It compares one automated model with economists supported by models, not humans working alone.
Michl predicts machines will outperform that team but still urges users to retain their own judgment.
Source: AI forecasts GDP as well as economists
IN: