Deep learning can beat other forecast methods – Bank of Korea research

Deep learning – an advanced form of artificial intelligence – can be more accurate in predicting outcomes, compared with conventional econometric approaches, according to research from Bank of Korea (BoK).

The research paper tested predictions of monthly exports from Korea and daily Korean won-US dollar exchange rates. It found that deep learning approaches produced better results even with the sorts of non-granular data sets that are normally used for conventional econometric models.

Central banks are showing increasing interest in using AI techniques – including deep learning and other types of machine learning – in research and monetary policymaking.

First developed in the field of computer science, machine learning algorithms train themselves through learning underlying patterns in the data. That means they can sometimes be wrong-footed if the pattern changes, but economists are exploring ways of overcoming this issue.

The BoK is among the central banks that have dedicated teams to explore use of AI techniques. Earlier in June, the Korean central bank said it would set up a digital innovation division to conduct research on digital currency and AI.

The BoK also plans to establish a new economic statistics system by the end of 2022, designed to take advantage of big data and machine leaning, in a bid to make forecasts more efficient and accurate.

Machine learning could help produce better economic growth forecasts amid high uncertainty due to the Covid-19 pandemic, the BoK says.

“Machine learning approaches prevail even with non-granular data which is used in conventional econometrics methods,” Soohyon Kim, economist with the BoK, writes in the deep learning paper. “This answer leads to a new era of econometrics and statistical methods, in which we can take advantage of advanced technologies of computing and processing micro-level data that used to be isolated from conventional econometrics.”

Early stages

The major hurdle in applying deep learning in economic analysis and forecasts is the data quality and volumes, the paper says.

“Unlike data that is dealt with in engineering, such as images, economic data has specific features, low frequency and a lot of noise,” Kim writes.  

The paper uses historical data to make short-term forecasts. It then compares the results of deep learning-based forecasts and traditional econometrics models.

Deep learning produces results with a noticeably narrower margin of error compared with the traditional method. The margin of error is especially small in the case of exchange rate forecasts, the paper says.

While the results are encouraging, the study is no more than a prototype, as it only uses historical data from one category, the author says. “Other than the examples presented in this paper, there are a lot of data we can exploit if we adopt deep learning more broadly.”

To train better models, central banks can use a broader range of data sources, such as real-time data from financial markets, and text data, for instance from newspapers.

“We hope that this study, as a starting point, demonstrates that the deep learning approach can be an alternative toolbox in addition to conventional econometrics,” the paper says.

A Federal Reserve study also explores applications of machine learning. When supplied with diverse and complex data, the machine learning model can outperform simpler time-series models, Aaron Smalter, a senior data scientist at the Kansas Fed, writes in the paper. In this case, machine learning models also outperform consensus estimates by professional forecasters.

In particular, a machine learning model can identify turning points in the unemployment rate earlier than competing methods, the Fed paper says. The results suggest machine learning models can provide forecasters with more guidance about cyclical fluctuations than a consensus or autoregressive forecast.