Rate Brief ENDE

Czech central bank's Michl: a lone AI model nowcasts GDP as well as staff

Czech National Bank governor Aleš Michl gave a lecture at Kyoto University on 13 November, published by the BIS on 19 November, in 33 short points. Most of it retold the bank's disinflation: inflation of 17.5% when he took office in mid-2022, a promise of close to 2% within two years, a strategy of keeping the key rate above inflation for longer and a strong koruna without intervention, and inflation at 2% in February and March 2024. He said the latest CPI reading is 2.4% and the key rate 4%.

The part that gave the lecture its title concerns forecasting. During COVID the bank built a weekly activity index from fast data such as electricity use and lorry traffic, and since 2023 it has used machine learning to weigh the dozens of indicators that arrive before GDP. Michl described a live contest, run since 2023 in real time rather than on past data, for the GDP forecast of the current quarter from the same data on the same day: on one side all the bank's models plus its economists' judgement, on the other a single Random Forest with no human involvement. The economists' average error was 0.21 percentage points; the machine's 0.22.

Michl summed up the bank's approach as "conservative in monetary policy, innovative in everything else". He listed instant payments launched in 2025, about $180 billion of reserves with the equity share doubled to 26% and more gold, a lab testing AI and tokenisation, a $1 million test portfolio of bitcoin, dollar stablecoins and a tokenised deposit, and a first test bond on a distributed ledger. AI, he said, already helps write and retrain the models, while "the judge remains human – but I don't know for how long".

Czech central bank's Michl: a lone AI model nowcasts GDP as well as staff
Czech central bank's Michl: a lone AI model nowcasts GDP as well as staff — Rate Brief

Why it matters

A gap of one hundredth of a percentage point suggests that, for short-term GDP estimates, staff judgement adds little over a well-trained model. The comparison is the bank's own and the details are in a CNB blog; it says nothing yet about forecasts further ahead, where policy decisions are actually made.

Written by Victoria Shinder.