Explainable AI for Small-Sample Macroeconomic Forecasting: A Structural Analysis of the Bangladeshi Economy

Abstract Gross Domestic Product Growth (GDPG) is an essential indicator of the economic development of a state. The GDPG is understood as the annual percentage change in the value added of goods and services provided within an economy. This paper deal with the problem of GDPG forecasting for volatile emerging economies due to structural instability and small sample sizes leading to high variance in both econometrics and machine learning methods. The XAI technique is proposed to forecast GDPG of Bangladesh by using the dataset from the World Bank covering years from 1991 to 2024. After the necessary econometric pre-processing to ensure stationarity and eliminate multicollinearity, five machine learning models were tested. The model with the highest level of regularisation is L2-regularised XGBoost algorithm, providing the best out-of-sample forecast (R-squared = 0.6245). The introduction of SHAP allowed for the automatic global dimensionality reduction, hence the detection of spurious variables. The model revealed structural asymmetries like short-term dilution of capital-labor and lagged limitations caused by the volatility of the real interest rate. In general, it could be stated that the correct calibration of the model made possible by the use of the proposed technique could provide accurate macroeconomic information for making decisions.