딥러닝 기반 시계열 예측 기법을 활용한 국제회의 수요에 관한 연구
A Study on the Demand for Convention using Deep Learning-based Time Series Forecasting
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This study forecasted and analyzed the number of international conferences held using three time series forecasting models: SARIMAX, GRU, and LSTM. In monthly forecasts, all three models faithfully reproduced the overall trend of actual values. GRU and LSTM, in particular, effectively reflected the long-term growth trend of GDP, an exogenous variable, demonstrating a gentle upward trend. In terms of annual aggregates, SARIMAX and LSTM showed similar trends and slopes, demonstrating a balanced reflection of seasonality and long-term trends. GRU produced slightly higher forecasts, but maintained a stable overall upward trend. This trend suggests that the long-term economic recovery effect was consistently reflected in the forecasts by incorporating exogenous variables. Although the training data in this study was limited to four years (48 months), which limited the seasonality learning of deep learning-based models, all three models achieved a certain level of forecast accuracy. The small performance gap between statistical and deep learning-based models suggests that both approaches can be practically applied, depending on the data characteristics and objectives.
Abstract
This study forecasted and analyzed the number of international conferences held using three time series forecasting models: SARIMAX, GRU, and LSTM. In monthly forecasts, all three models faithfully reproduced the overall trend of actual values. GRU and LSTM, in particular, effectively reflected the long-term growth trend of GDP, an exogenous variable, demonstrating a gentle upward trend. In terms of annual aggregates, SARIMAX and LSTM showed similar trends and slopes, demonstrating a balanced reflection of seasonality and long-term trends. GRU produced slightly higher forecasts, but maintained a stable overall upward trend. This trend suggests that the long-term economic recovery effect was consistently reflected in the forecasts by incorporating exogenous variables. Although the training data in this study was limited to four years (48 months), which limited the seasonality learning of deep learning-based models, all three models achieved a certain level of forecast accuracy. The small performance gap between statistical and deep learning-based models suggests that both approaches can be practically applied, depending on the data characteristics and objectives.