考量參考價格之需求參數估計與動態訂價

Abstract

本研究將深度學習技術應用於需求參數估計與動態訂價，訓練出能在未知市場環境做精準估計並即時進行動態決策優化的模型。 考量參考價格的動態訂價決策對販售時效性商品的業者來說至關重要。市場環境中的消費者平均價值常難以量化且取得不易，若無法有效率地估計以掌握消費者需求輪廓，將影響訂價決策並造成獲利損失。本研究提出之方法與開發之模型將解決上述困境。 本研究探討環境為販售單一時效性商品的市場，目標是令預期利潤最大化——應用多項羅吉特模型(Multinomial Logit Model，MNL)將消費者平均價值與參考價格納入動態訂價狀態參數，求解所有狀態集合下的最佳定價策略，以達預期利潤最大化之目的；其次，使用大量歷史銷售資料作為訓練樣本，開發出可在未知市場環境中精準估計消費者平均價值之長短期記憶神經網路(Long Short-Term Momery，LSTM)；最後，藉由自適應訓練(Adaptive Training)逐步提升LSTM估計效能及訂價決策成效，令模型得利用真實銷售資料來適應當前市場環境，並透過離散事件模擬來驗證模型於未知環境中的估計準確度與通用性。實驗結果證實，本研究開發之模型在未知市場環境也具備良好效能，可供實務應用參考。

This research applies deep learning method in demand parameters estimation and dynamic pricing to train a model that can make accurate estimation in unknown market environments and conduct decision optimization in real time. Good dynamic pricing strategy considering reference prices are crucial to sellers with short selling season. Mean of customers’ valuation are usually difficult to quantify and estimate. Being unable to estimate demand parameters efficiently will lead to bad pricing strategy and considerable profit loss. Our study proposes a model to solve the difficulties mentioned above. We consider a dynamic pricing problem in a market selling single perishable product with the goal of maximizing expected profits. We take mean of customers’ valuation and reference prices into the consideration of dynamic pricing with the application of multinomial logit model, solving the optimal pricing strategies under all possible states. Then we use a large amount of historical sales data to develop a long short-term memory neural network which can precisely estimate mean of customers’ valuation in unknown market environments. Finally, we apply adaptive training to improve the performance of LSTM’s estimation and pricing strategy, let model be self-adaptive to current market environments by utilizing real-time sales data, then use discrete events simulation to verify our model’s accuracy and generalization under unknown environments. The result indicates that our model performs well in a variety of unknown market environments.