Earticle

본 연구는 우리나라 수출 분야 산업의 경쟁력을 나타내는 부가가치율을 다양한 머신러닝 기법을 활용하여 예측하였다. 아울러, 예측의 정확성 및 안정성을 높이기 위하여 머신러닝 기법 예측값들에 예측조합 기법을 적용하였다. 특히, 본 연구는 산업별 부가가치율에 영향을 주는 다양한 변수를 고려하기 위하여 재귀적특성제거 방법을 사용하여 주요 변수를 선별한 후 머신러닝 기법에 적용함으로써 예측과정의 효율성을 높였다. 분석결과, 예측조합 방법에 따른 예측값은 머신러닝 기법 예측값들보다 실제의 산업 부가가치율에 근접한 것으로 나타났다. 또한, 머신러닝 기법의 예측값들이 큰 변동성을 보이는 것과 달리 예측조합 기법은 안정적인 예측값을 나타내었다.

This study predicts the ratio of added value, which represents the competitiveness of export industries in South Korea, using various machine learning techniques. To enhance the accuracy and stability of prediction, forecast combination technique was applied to predicted values of machine learning techniques. In particular, this study improved the efficiency of the prediction process by selecting key variables out of many variables using recursive feature elimination method and applying them to machine learning techniques. As a result, it was found that the predicted value by the forecast combination method was closer to the actual value than the predicted values of the machine learning techniques. In addition, the forecast combination method showed stable prediction results unlike volatile predicted values by machine learning techniques.

To effectively predict auto sales and improve the competitiveness of automotive enterprise, the characteristics of actual auto sales were analyzed, owing to the seasonal fluctuations and the nonlinearity of monthly sales, the combination forecasting model based on seasonal Index and RBF neural network was proposed. The weights of the two single models were computed using mean absolute percentage error and the sum of square error respectively, the result shows that mean absolute percentage error is more effective. Finally, the prediction accuracy of different models was compared based on the criteria of MAPE and RMSE, and the effectiveness of the method was proved, the proposed model can take advantage of the strengths of the two single models, the results indicate that the combination forecasting model suitable for auto sales has high prediction accuracy, which can provide a certain reference to auto sales forecasting.

To effectively predict cigarette sales and improve the competitiveness of tobacco business enterprises, the characteristics of actual cigarette sales were detailed analyzed. Due to the long-term growth trends, seasonal fluctuations and the nonlinearity of monthly sales, we established three single forecasting models, which are Exponential Smoothing (ES), Seasonal Decomposition (SD) and Radial Basis Function (RBF) neural network. After obtaining the predicted value of three single models, the combination forecasting model was proposed. The weights of the three single models were computed using Mean Absolute Error and the mean relative error respectively, the result shows that relative error is more effective. A dynamic weight combination forecasting method based on RBF is proposed and compared with fixed weight method. Finally, the prediction accuracy of different models was compared based on the criteria of MAPE and RMSE, and the effectiveness of the combination method was proved, the proposed model can take advantage of the strengths of the three single models, the results indicate that the combination forecasting model suitable for cigarette sales has higher prediction accuracy. In some cases, the prediction accuracy of the fixed weight combination model is better than the dynamic weight combination model. The results can provide a certain reference to cigarette sales forecasting.

The combination forecasting model IOWGA-EMD-ARMA-WNN is proposed in this paper. The randomness, periodicity and tendency of the original data are showed by EMD decomposition in EMD-ARMA model. WNN combines the advantages of wavelet analysis and BP neural network and improves the learning efficiency and forecasting accuracy. The weight of combination model is decided by forecasting precision of EMD-ARMA model and WNN model based on IOWGA method. At last, the IOWGA-EMD-ARMA-WNN model is used to forecast monthly inboard tourism demand of China and the results show that the proposed combination model has better performance on forecasting accuracy compared with the other models.

High-speed railway subgrade settlement forecasting is an important part to ensure ride comfort and stability of the high-speed railway line. Based on the measured settlement data of Lan-Xin Railway DK649+500D-K650+000 section, three kinds of single forecast methods, including Hyperbolic method, Three-point method and Asaoka method are adopted for the high-speed railway subgrade settlement forecasting. The least squares support vector machine in adopted for its unique advantages in solving small sample data and nonlinear data. Combination forecast model is established by setting different weights of the single forecast methods. Simulation results show that the systematic errors in the Combination forecast model are less than that in the single forecast methods. Compared with the single forecast methods, the Combination forecast model has higher forecasting accuracy and better applicability.

최근 들어 경제전망의 예측 오차를 줄이기 위한 방안으로 예측조합에 의한 전망(forecast combination)과 밀도함수에 의한 전망(density forecast)이 주목받고 있다. 예측조합에 의한 전망은 하나의 모형에 의한 전망이 아니라 다양한 특성을 가진 여러 예측모형들을 조합하여 전망치를 산출함으로써 예측 오차를 줄이는 방법이다. 밀도함수에 의한전망은 점 예측(point forecast)과 달리 전망치에 대한 확률밀도함수(probability density function)를 제시하는 방법이다. 본고에서는 우리나라의 소비자물가 상승률을 대상으로 예측조합 및 밀도함수를 이용하여 전망을 수행하고 그 성과를 분석하였다. 예측조합에 의한 전망의 경우 다양한 모형의조합에 따른 위험분산효과 등으로 기준모형인 자기회귀모형보다 우월한 예측력을 보였으며, 아울러 더 많은 수의 모형들이 조합될수록 보다 안정적인 예측력을 보였다. 한편 밀도함수에 의한 전망의 경우 소비자물가 상승률 전망의 방향성을 제시해주는 장점이 있었다. 또한 밀도함수 역시 조합을 통해 보다 정확한 물가 상승률 전망을 기대해 볼 수 있다. 본고의 연구 결과에 비추어 볼 때 향후 모형들의 조합방법과 모형별 가중치 부여방법에 대한 연구, 혼합주기모형 개발 등이 진전될 경우 예측조합이나 밀도함수에 의한 소비자물가 상승률 전망의 예측력을 더욱 향상시킬 수 있을 것으로 기대된다.

Forecast combinations and density forecast have frequently been found in empirical research to produce better prediction performance on average than methods based on the best single model. Density forecast an estimate of the probability distribution of the possible future values of that variable has received attention in the forecast literature. This paper combines point forecast and density forecast to predict Korean CPI inflation and compares the performance of each forecast with various models including factor models, shrinkage models, and bayesian model averaging. We find that the more models included in point forecast combinations leads to the better performance of the combinations than the benchmark autoregressive model, regardless of the independent performance of a single model. We also find that combinations of more models provide a result robust to sample periods. Density forecasts and their combinations present the direction of future inflation and predictive densities. We expect that forecast combination and density forecast can provide better performance with more disciplines, for example, combining more various models and mixing different frequency data models.

본 연구에서는 머신러닝(Machine Learning, ML)과 딥러닝(Deep Learning, DL) 모델을 앙상블(Ensemble)하여 어떠한 주가 예측 방법이 우수한지에 대한 연구를 하고자 한다. 연구에 사용된 모델은 하이퍼파라미터(Hyperparameter) 조정을 통하여 최적의 결과를 출력한다. 앙상블 방법은 머신러닝과 딥러닝 모델의 앙상블, 머신러닝 모델의 앙상블, 딥러닝 모델의 앙상블이다. 세 가지 방법으로 얻은 결과를 평균 제곱근 오차(Root Mean Squared Error, RMSE)로 비교 분석하여 최적의 방법을 찾고자 한다. 제안한 방법은 주가 예측 연구의 시간과 비용을 절약하고, 최적 성능 모델 판별에 도움이 될 수 있다고 사료된다.

A new technique, namely the combination of satellite and trajectory analysis (CSTA), for exploring the spatio-temporal distribution information of volcanic ash plume (VAP) from volcanic eruption. CSTA uses the satellite derived ash property data and a matching forward-trajectories, which can generate airmass history pattern for specific VAP. In detail, VAP properties such as ash mask, aerosol optical thickness at 11 ${\mu}m$ ($AOT_{11}$), ash layer height, and effective radius from the Moderate Resolution Imaging Spectro-radiometer (MODIS) satellite were retrieved, and used to estimate the possibility of the ash forecasting in local atmosphere near volcano. The use of CSTA for Iceland's Eyjafjallaj$\ddot{o}$kull volcano erupted in May 2010 reveals remarkable spatial coherence for some VAP source-transport pattern. The CSTA forecasted points of VAP are consistent with the area of MODIS retrieved VAP. The success rate of the 24 hour VAP forecast result was about 77.8% in this study. Finally, the use of CSTA could provide promising results for VAP monitoring and forecasting by satellite observation data and verification with long term measurement dataset.

본 연구는 충남 경제에 대한 사례 분석을 통해 자기회귀 프로빗 모형에 기초한 확률예측결합이 지역경제 수축기 예측력을 높이는 데 큰 도움이 된다는 점을 보인다. 주요 결과는 다음과 같다. 첫째, 본 연구의 방법은 실시간 예측 상황에서 현 시점 경기국면을 최대 93.6%, 3개월 후 경기국면을 최대 78.4%의 비율로 정확히 예측한다. 둘째, 프로빗 모형 중에서는 경기국면의 지속성을 고려하는 자기회귀 모형이 그러한 특성을 고려하지 않는 정태적 모형보다 예측력이 높다. 셋째, 확률예측결합은 예측의 정확성 뿐 아니라 안정성을 제고하는 데도 도움이 된다. 예측결합 방법 중에서는 각 개별 모형의 과거 예측력에 근거해 가중치를 달리하는 방법이 단순 평균 방법보다 우수한 성과를 산출한다.

This study examines the performance of autoregressive probit models and forecast combination methods for predicting economic downturns in Chungnam, a rural province of South Korea. Its results of out-of-sample forecasting exercise for the period 2007:M1-2017:M5 are summarized as follows. First, the methods considered in this paper can greatly enhance the real-time forecasting ability for the occurrence of regional recessions. In particular, they can accurately predict the current and three month ahead business-cycle phases at a rate of up to 93.6% and 78.4%, respectively. Second, among the probit models, the autoregressive model that takes into account the persistence of business-cycle phases provides more accurate forecasts than the static model that does not consider such empirical features. Third, the forecast combination methods improve significantly not only the forecast accuracy but also the stability of predictive ability. Among those combination methods, the different weighting schemes based on the past forecasting performance of each predictor variable yields better results than the simple averaging method.

본 연구는 우리나라 소비자전망지수의 세분화된 자료에 다양한 예측기법을 적용하여 그 중 어떠한 방법이 가장 우수한 성과를 보이는지를 분석한다. 예측대상 변수로는 산출, 소비, 고용, 물가, 금리 변수를 고려한다. 정보변수로는 다양한 기준에 따라 26개 그룹으로 세분화된 소비자전망지수를 고려한다. 예측기법으로는 전통적인 ARDL 방법뿐만 아니라, 단순 예측결합방법, 할인MSFE 가중평균법, 베이지안 모형평균법, 정보기준 가중평균법, 클러스터 결합방법, 주성분기법 등을 고려한다. 분석 결과 소비자전망지수가 거시경제 변수들을 예측하는 데 도움이 될 수 있으며, 클러스터 기법이 해당 지표의 활용도를 높이는 최선의 방법인 것으로 나타났다. 아울러, 클러스터 기법 중에서는 클러스터별 가중치를 OLS로 추정하는 방법과 이를 베이지안 기법으로 수축하는 방법이 가장 우수한 성과를 보이는 것으로 나타났다.

Different consumers may form different expectations about future economic conditions. Therefore we need to combine individual forecasts rather than relying a single forecast from a particular consumer group. In this paper we investigate the relative performances of several combination methods in predicting economic variables such as output, consumption spending, employment, unemployment rate, price level, and interest rate, using disagregated data for the consumer expectations. In addition to the principal component method, we consider forecast combination methods based on simple averaging, discount MSFE, Bayesian shrinkage, information creteria, and clustering. We find that consumers' expectation indexes are useful to predict most of the economic variables we consider. We also find that the combination methods based on clustering outperform other forecasting methods, especially when the weight of each cluster is determined using ordinary least squares estimation or Bayesian shrinkage techniques.

본고는 다중 공적분 시스템 하에서 임의보형 모형과 공적분 모형의 최적 선형결합 예측기법(OBLP)을 이용하여 한국의 GDP와 소비를 예측하고 그 예측 성과를 임의보형 모형, 공적분 모형 및 VAR 모형과 비교한다. 여기서 공적분을 구성하는 펀더멘털로는 통화·재정정책 및 대외충격을 반영하여 이자율 기간 스프레드, M1, 무역수지 및 정부 지출을 우선 고려하였다. 제안된 모형의 표본 외 예측 성능을 분석하기 위해 1987년 1분기부터 2017년 4분기까지의 자료를 사용하여 2018년 1분기부터 2022년 4분기까지 각 분기의 한국 GDP와 소비를 예측하였다. 그 결과, 모든 경우의 예측 오차의 절대값 기준에서 OBLP 모형이 임의보행 모형을 능가하며, 10분기 이후 예측 결과에서 공적분 모형보다 우수한 것으로 나타났다. 다만 10분기 이전의 경우 OBLP와 공적분 모형간의 예측성과는 우열이 혼재하는 양상을 보였다. VAR(2) 모형은 1분기 앞을 제외하고는 여타 모형에 비해 매우 큰 예측 오차를 나타냈다. 이런 예측 우수성을 감안할 때 정부나 민간기관의 장단기 예측에서 OBLP 모형을 함께 고려하는 것도 매우 유용할 것으로 판단된다.

This paper forecasts Korean GDP and consumption using the optimal linear combination forecasting technique (OBLP) of random walk and cointegration models under a multiple cointegration system and compares its forecasting performance with random walk, cointegration, and VAR models. We first consider the interest rate term spread, M1, trade balance, and government expenditure as the fundamentals of the cointegration model to reflect monetary and fiscal policies and external shocks. To analyze the out-of-sample forecasting performance of the proposed model, we use data from 1987Q1 to 2017Q4 to forecast Korean GDP and consumption for each quarter from 2018Q1 to 2022Q4. The results show that the OBLP model outperforms the random walk model in terms of the absolute value of forecast errors in all cases, and outperforms the cointegration model in forecasting results after 10 quarters. However, prior to the 10th quarter, the predictive performance of the OBLP and cointegration models is mixed. The VAR(2) model has a much larger forecast error than the other models except for the first quarter. Given this forecasting superiority, it would be very useful to consider the OBLP model in the short- and long-term forecasts of the government, the Bank of Korea, or private organizations.