The Long Short-Term Memory Algorithm and the Autoregressive Integrated Moving Average Approach in Business Tendency Survey

Abstract

This study investigates the effectiveness of machine learning models in forecasting construction indicators derived from Business Tendency Survey data. Specifically, we compare the performance of traditional statistical models such as the autoregressive integrated moving average (ARIMA) with long shortterm memory (LSTM) networks and hybrid approaches combining both. Using a range of economic variables – including sector and economic evaluations, production, financial situation, investments, and sentiment indicator (IRGBUD) - we evaluate model accuracy across testing dataset and rolling forecast strategy to assess consistency over time. Results demonstrate that while LSTM networks capture non-linear dependencies and temporal patterns, ARIMA-based models consistently outperforms LSTM in scenarios involving seasonal and cyclical structures. The findings highlight that the choice of model should align with the nature of the time series, particularly in relation to seasonality, volatility, and trend dynamics. This work offers practical implications for improving economic forecasting with machine learning in survey-based environments.