논문 상세보기
AI Framework for Strategic Decision Support in the Shipping Market through the Integration of Transformer and Kalman Filtering KCI 등재
Transformer와 Kalman Filter 결합을 통한 해운시장의 전략적 의사결정 지원 인공지능 프레임워크
- 언어KOR
- URLhttps://db.koreascholar.com/Article/Detail/449276
pp.57-70
Freight-rate forecasting in the VLCC TD3C market remains challenged by abrupt regime shifts, pronounced volatility, and heterogeneity in real-time signals from oil prices, seaborne trade, vessel operations, and macroeconomic factors; these directly impact freight planning and chartering. This study presents a daily multivariate dataset with 4,267 samples covering 2014-02-01 to 2025-10-08, integrating crude benchmarks, fuel spreads, refinery margins, port congestion, inventory levels by region, plus detailed AIS-derived VLCC activity, speed, and operation states, scaled and split 80/10/10 for training, validation, and testing. The proposed framework combines a PyTorch Transformer—optimized using Optuna for d_model=128, 9 layers, 8 heads, a 14-day input window, and 5-day output—with Monte Carlo Dropout for uncertainty quantification. Diagnosis uses differential entropy and coefficient-of-variation to verify convergence with 90 separate runs, while a Kalman filter (Q=0.001, R=0.01) smooths the forecast trajectory and enhances temporal reliability. Experimental results show baseline Transformer achieves average MAE 5,259.4, MAPE 13.10%, and R²=0.74 across 1-5 day horizons, with volatility quality metrics declining at longer leads. Applying the Kalman filter reduces errors to MAE 4,326.1, MAPE 10.6%, and raises R² to 0.83; timing and extremity components of volatility quality scores are strengthened, providing a more robust basis for operational decisions. Monte Carlo backtesting for 82 Korean VLCCs over 598 trades finds the Kalman-smoothed strategy earns $108.5M (88.9% win rate, Sharpe ratio 0.83), substantially outperforming raw Transformer ($32.9M, 60.5%, 0.24) and random selection (near zero, 49.3%, 0.005). These results highlight the clear economic value added by calibrating uncertainty and post-processing forecasts, transforming predictive reliability into real-world freight portfolio improvement in the tanker market.
2. 관련 연구
2.1 해운 운임 예측 연구
2.2 Transformer 모델
2.3 Kalman Filter의 시계열 예측 응용
3. 데이터 수집 및 전처리
4. 실험 결과 및 분석
4.1 모델 개요와 시스템 아키텍쳐
4.2 변수 및 하이퍼파라미터 및 최적화
4.3 Monte Carlo 앙상블을 통한 시계열 예측신뢰도 및 성능 향상
4.4 Transformer, ARIMAX, LSTM 성능비교
4.5 Transformer 성능 평가와 한계
4.6 Kalman Filter로 강화한 Transformer 예측
5. 백테스트를 통한 경제적 효과 분석
5.1 현실 제약을 반영한 VLCC 운영 백테스트
5.2 예측전략 별 성과 비교 및 의사결정 분석
5.3 프레임워크 효과와 운영․정책적 시사점
6. 결 론
Acknowledgement
References
