AI-enhanced reserve management for China during economic crises: an integrated machine learning and buffer inventory approach

This study aims to explore China's optimal reserve strategy during crises. To this end, this study developed an AI-based inventory management framework for China, which combines random forest/gradient boosting tree (GBDT) ensemble learning with buffer inventory safety stock rules, and validated it using data from the International Monetary Fund (IMF), the State Administration of Foreign Exchange (SAFE), and Bloomberg from 2000 to 2023. Compared to traditional linear models, the integrated model reduces the mean squared prediction error of optimal reserves by 74% (3.21 vs. 12.34), and compared to using only the random forest model, it reduces the error by 44% (3.21 vs. 5.78). Empirical elasticity analysis shows that a 1% increase in GDP growth rate leads to a 0.56% increase in optimal reserves (t=12.34); conversely, a 1% increase in exchange rate volatility results in a 0.32% decrease in optimal reserves (t=–7.65). Under a crisis scenario (trade balance of-$50 billion, exchange rate volatility of 6%), the framework recommends immediately adjusting the reserve composition, increasing the gold ratio from 10% to 25% and the strategic materials ratio from 5% to 15%, which is expected to reduce the anticipated shortfall cost by $3.2 billion annually. Through robustness tests using 10 random subsamples and extreme value cleaning, coefficient deviations are all controlled within ±7%. The framework demonstrates significant stability in terms of parameter adjustments and sample variations. The study proposes policy recommendations including improving reserve decision-making, strengthening buffer stock mechanisms, promoting asset diversification, and establishing intelligent management systems. These recommendations provide a scientific framework for reserve operations during crises.

Text 125921.pdf - Published Version Restricted to Repository staff only Download (643kB) Official URL or Download Paper: https://www.aasmr.org/liss/Vol.12/No.5/Vol.12.No.5...

Actions (login required)

View Item