Analysis of foreign trade using artificial intelligence methods: The case of Türkiye (2000-2022)

In this study, Türkiye's export data to 152 countries between 2000 and 2022 was utilized to generate export forecasts for the 2023–2027 period. The primary objective of the study is to comparatively evaluate the forecasting performance of two artificial intelligence-based models Long Short-Term Memory (LSTM) and Support Vector Regression (SVR) in predicting foreign trade. Initially, the period of 2018–2022 was set aside as the test dataset, and in-sample forecast performance was evaluated for both models. Subsequently, out-of-sample forecasting was conducted for the 2023–2027 period, which includes years the models had not previously encountered. Since real-world values of independent variables (such as GDP, real effective exchange rate, and Linder index) were unavailable for 2023–2027, these variables were projected using the exponential smoothing method based on historical time series trends and were used as inputs to the forecasting models. According to the modeling results, both LSTM and SVR demonstrated high predictive accuracy for the 2018–2022 test period, with R² scores calculated as 86.71% for LSTM and 83.61% for SVR. In the out-of-sample prediction phase (2023–2024), an improvement in model accuracy was observed, with R² scores rising to 94.58% for LSTM and 95.02% for SVR. These findings indicate a strong statistical fit when using transformed data. The LSTM model exhibited an upward trend in export volume forecasts for future years but also showed increased variance and uncertainty over time. In contrast, the SVR model provided more stable and controlled forecasts in the short term, although notable deviations were observed for some countries. Model performance was evaluated using classical regression metrics such as RMSE, MAE, and R². Furthermore, countries were classified into three accuracy categories (accurate, moderately accurate, and deviated) based on the ratio of predicted to actual values, which helped identify where each model performed more effectively. In conclusion, the study demonstrates that both LSTM and SVR can be effectively applied for foreign trade forecasting. However, model selection should consider factors such as data structure, forecast horizon, and economic context. By presenting a comparative analysis of LSTM and SVR in this domain, the study provides a novel contribution to the national literature on artificial intelligence applications in trade forecasting.