結合長短期記憶和大型語言模型之混合型銷售預測模式

摘 要

個案公司主要供應國際汽車大廠的各式扣件，目前都是接到訂單後以經驗法則安排後續作業，但面臨原物料採購前置時間長且需求變動大的情況下，常出現待料停機和延遲交貨的問題。對此，本研究提出一結合量化與質性分析的混合式銷售預測模式，協助管理者提前掌握銷售變動，制定更韌性的原物料庫存計畫。量化預測方面使用ERP歷史訂單，以ARIMA驗證時間序列特性，再訓練建構長短期記憶模型（LSTM）提供銷售預測值，結果顯示排除疫情期間資料的判定係數 (R²) 達到 0.851。質性分析方面則是收集775篇汽車市場景氣報導，微調BERT語言模型使其能辨識文章的景氣分類，預測時將當月多篇景氣文章的分類彙整呈現，提供管理者綜合判斷後，適度調整銷售預測值，藉以彌補僅靠時間序列變動軌跡預測時未反應外部變化的風險。本研究混合型方法能更全面地考慮經營環境因素，可有疾改善銷售預測準確性，提昇企業滿足訂單需求的能力。

Abstract

The case company primarily supplies various fasteners to international automotive manufacturers. The company currently relies on experience-based rules to schedule subsequent operations after receiving orders. However, given the long lead times for raw material procurement and high demand variability, the company frequently experiences machine downtime caused by material shortages and delayed deliveries. To address this issue, this study proposes a hybrid sales forecasting model that integrates quantitative and qualitative analyses. For the quantitative analysis, historical order data from the ERP system is used. The ARIMA model is employed to verify the time series characteristics, followed by the training and construction of a LSTM model to generate sales forecasts. Results indicate that the R² reaches 0.837 after excluding data from the pandemic period. For the qualitative analysis, 775 automotive market reports were collected, and the BERT language model was fine-tuned to identify the sentiment classification of the articles. During forecasting, the classifications of multiple market articles in the target month are aggregated and presented to managers, helping them to make comprehensive judgments and adjust the sales forecasts accordingly. This approach mitigates the risk of failing to reflect external changes when solely relying on time series trends for forecasting. The hybrid approach of this study offers a more comprehensive consideration of external environment factors, significantly improves sales forecast accuracy, and enhances the company's ability to meet customer order demands.