This study aims to explore predictive variables for one-period-ahead sales forecasting in manufacturing firms using corporate financial data and public environmental indicators and to compare the predictive performance of machine learning models. For this purpose, a panel of 376 KOSPI-listed manufacturing firms from 2010 to 2024 was used, yielding 5,640 firm-year prediction pairs. The analysis applied Linear Regression, Elastic Net, Random Forest, and Extreme Gradient Boosting (XGBoost), and used expanding-window external validation for input years 2021-2024, time-ordered inner validation for hyperparameter selection, conditional variable assessment after fixing period-t log sales, and ablation analysis of corporate financial data, public data, and their combinations. The main empirical results are as follows. Period-t log sales provided the strongest baseline information for period-t+1 log sales. After fixing period-t log sales, revenue per employee, total capital growth rate, value added ratio, the ratio of depreciation to total costs, and the ratio of labor costs to total costs reduced RMSE, although the reductions were small and most were not statistically significant after Holm adjustment. In the comparison using corporate financial data, Linear Regression achieved the lowest numerical prediction error, with a Mean Squared Error (MSE) of 0.1436, a Root Mean Squared Error (RMSE) of 0.3789, a Mean Absolute Error (MAE) of 0.1891, and an R² of 0.9524. Elastic Net produced an RMSE of 0.3826, XGBoost produced an RMSE of 0.3893, and Random Forest produced an RMSE of 0.3983, while pairwise loss tests did not establish the consistent superiority of Linear Regression over all alternative models. Public-data-only models produced negative R² values, and adding contemporaneous public indicators or industry-sensitive lagged indicators did not improve predictive performance. In the tuned XGBoost model, the corporate-financial-data specification produced a lower RMSE than the single-variable specification, and SHAP analysis identified period-t log sales as the most influential variable. These results indicate that firm-specific financial information provides additional predictive value beyond prior sales, whereas annual public indicators require a data structure that more directly captures industry- and firm-level exposure.

1. 서 론
2. 이론적 배경
    2.1 제조업 매출의 지속성과 기업 재무정보의예측기여
    2.2 외부 공공환경지표의 예측정보와 업종별 반응
    2.3 머신러닝 기반 기업성과 및 매출 예측 관련선행연구
3. 연구 방법
    3.1 분석자료와 변수 구성
    3.2 자료 전처리 및 예측쌍 구성
    3.3 분석설계와 자료구성별 비교
    3.4 시간순 중첩 검증
    3.5 예측모형과 하이퍼파라미터 탐색
    3.6 조건부 변인 선택과 공공데이터 증분기여 분석
    3.7 성능평가와 통계검정
4. 연구결과
    4.1 표본 특성과 기술통계
    4.2 변인별 예측성능과 누적 투입 결과
    4.3 기준매출 통제 후 조건부 변인 기여
    4.4 하이퍼파라미터 탐색에 따른 모형별예측성능과 차이검정
    4.5 자료구성별 공공데이터 증분기여
    4.6 외부검증 오차와 음의 R²
    4.7 XGBoost의 변수구성별 성능과 설명가능성분석
    4.8 Random Forest의 외삽오차 진단
5. 결 론
References

• Se-Hyeong Cho(Department of Industrial and Management Engineering, Incheon National University) | 조세형 (인천대학교 산업경영공학과) Corresponding author
• Woo-Sik Yoo(Department of Industrial and Management Engineering, Incheon National University) | 유우식 (인천대학교 산업경영공학과)