The Bank of Korea Uses Deep Learning for Exchange Rate and Export Forecasting Analysis... "Reduced Margin of Error"

BOK Economic Research
Analysis and Forecasting of Macroeconomic and Financial Variables Using Deep Learning

Prediction range (red) and actual exchange rate (black line) using Vector Error Correction Model (above) and Deep Learning (below). It can be seen that the prediction range is significantly narrowed when using Deep Learning. (Source: Bank of Korea)

[Asia Economy Reporter Eunbyeol Kim] The Bank of Korea, which established the Digital Innovation Office last month, has begun analyzing whether it is possible to predict macroeconomics and financial markets using 'Deep Learning,' a machine self-learning technology. Analyzing exports and exchange rates showed that using deep learning can reduce the margin of error.

On the 2nd, Kim Suhyun, head of the Bank of Korea's Digital Innovation Office, presented "Analysis and Prediction of Macroeconomic and Financial Variables Using Deep Learning," stating, "As a result of analysis using deep learning, the prediction error range for exports was significantly narrow, and for the won-dollar exchange rate, the deep learning prediction results that removed anomalies in the time series were superior in both predictive power and error range."

When analyzing exchange rates, using deep learning improved predictive power and significantly reduced the margin of error. Looking at the prediction performance comparison indicator RMSE (Root Mean Squared Error), the vector error correction model showed 6.82 won, while deep learning showed 5.26 won. The error range was also narrower with deep learning (5~40 won) compared to the model (30~40 won).

However, when unusual events occurred, such as the Canadian government's arrest of Meng Wanzhou, Huawei's vice-chairwoman, at the request of the United States, the Chinese foreign exchange authority's allowance of 'Pochi (破七, the yuan exchange rate exceeding 7 yuan per dollar),' and the cancellation crisis of the US-China trade negotiations, the margin of error increased.

Analyzing customs-based export data showed that both traditional analysis models (vector autoregression model) and deep learning had similar predictive power. However, in terms of the margin of error, while traditional models ranged around 15 to 20 billion dollars, using deep learning narrowed the error range to about 5 to 10 billion dollars.

Kim said, "Currently, we are internally setting deadlines step-by-step to forecast and predict various exports, exchange rates, and gross domestic product (GDP), and plan to apply them gradually," adding, "We are also preparing to apply algorithms to GDP."

He also explained, "If deep learning can be applied to macroeconomic and financial variables, it will be possible to complementarily utilize various structured and unstructured data, which will help improve economic forecasting methodologies." However, he added that further review is needed before officially releasing forecasts.

Forecasting using deep learning is expected to be more efficient when unexpected crises such as the novel coronavirus infection (COVID-19) occur. Traditional economic forecasting models predict based on past data, but deep learning can acquire various real-time information such as text and images to make forecasts.

Last month, the Bank of Korea established the 'Digital Innovation Office' through a second-half organizational restructuring. The Digital Innovation Office, installed directly under the Deputy Governor in charge, conducts research on policy implementation and internal management application plans for new digital technologies such as artificial intelligence (AI) and big data, and establishes data governance policies and strategies at the bank-wide level. It also leads the expansion of the latest digital infrastructure such as the research platform (BReiT) and Data Lake.

This content was produced with the assistance of AI translation services.