競爭動態：剖析企業競爭行動與預測未來行動組合

Abstract

在變動迅速的商業世界中，企業之間的互動與關聯顯得益發複雜，以至於確實了解競爭對手並精準地預判其可能之競爭行爲即顯得至關緊要。這些競爭行爲相當常見，舉凡價格戰、行銷企劃、訴訟等皆屬之；而當一家企業發起這樣的行爲時，其競爭對手也會因此迅捷地反應，調整自己的策略，以求不再競爭行列中脫隊。是故，若企業能夠有效地分析當下的競爭態勢，乃至於精確地預測出其競爭對手的潛在行爲，將大大地有助益。然因搜集競爭資料並進行全面分析會引發之成本甚鉅，要做到有效分析、精確預測並不容易。 本研究從「競爭動態」的角度出發，試分析企業所面臨之競爭態勢。競爭動態是一個以競爭行爲爲主體、描繪在策略層面上，競業之間如何反覆利用不同行爲進行競爭的框架。我們看到過去的文獻已針對競爭動態進行一定的研究，包含進行個案探討、提出解釋性模型等，然而我們亦發現，預測競爭行爲本身之相關文獻甚少，實爲該領域有待深掘之處。我們因此在本研究中提出一個涵蓋競爭行爲之分類、分群與預測的端到端的分析流程。我們搜羅了大量新聞文件，建構了一自動分類器來將文件依所提及之競爭行爲分類、整合，藉此爲每家標的企業梳理出其過往的歷史競爭行爲。接著，我們分別利用二種不同的方法——依文件共同提及或是共同歷史行爲——爲標的企業識別出其最主要的競爭企業以便爲標的企業建構其競爭態勢。根據標的企業的競爭態勢，我們提出了一個以閘門循環單元（GRU）和注意力機制（Attention）爲基底的時間序列預測模型，該模型被用來預測標的企業在下一個時間點所會進行之不同競爭行爲之次數。 此研究主要集中在航空業，其不但以高度競爭聞名，資料也相對容易取得，過去研究和其使用的公開新聞資料的量體就證明了這一點。我們使用與航空公司相關的新聞文件資料進行實驗，並證明建構以深度學習爲基底的時間序列預測模型會表現得比傳統統計模型更佳；我們也探討了用不同方法識別競爭對手如何影響預測結果。這項研究不僅能協助產業內的企業進行決策，我們的預測結果亦對投資者、第三方諮詢行業提供資訊，成爲其決策過程的一部分。

In a fast-paced business world, interactions between firms are increasingly complex, making the understanding and prediction of competitors’ actions critical for strategic decision-making. Competitive actions often include price wars, marketing campaigns, litigations, and more. When such actions are initiated, it is crucial for rival firms to react swiftly and update their strategies to maintain their market positions. Therefore, firms benefit significantly from effectively analyzing the current competitive landscape and predicting future actions to gain an advantage over their competitors. However, performing such predictions is challenging due to the difficulty in data collection and efficient comprehension. In this study, we analyze the competitive landscape through the lens of competitive dynamics, an action-based perspective that investigates how rival firms compete through specific actions within their strategic contexts. Although previous studies in this field have proposed case studies and explanatory models to estimate the volume and complexity of competitive actions, directly predicting the occurrences of specific competitive actions has not been extensively explored. Our study presents an end-to-end pipeline to profile and forecast firms’ competitive actions using advanced machine learning techniques. By leveraging a large dataset of news articles, we built an auto-profiler that annotates articles and consolidates them to form competitive events for each firm. We then identify the main competitors for a firm based on criteria such as co-mentioning in articles and the frequency of performing matching actions. Finally, we constructed a time series prediction model incorporating Gated Recurrent Unit (GRU) and Attention mechanisms capable of predicting the frequency of certain competitive actions in the future. Our study focuses mainly on the airline industry, which is known for its high competitiveness and the relative ease of obtaining data, as evidenced by previous research and the frequency of public news data. Using a dataset of airline-related news articles, our experiments demonstrate that deep learning-based time series prediction models outperform traditional statistical models. We also examine how different mechanisms for competitor identification affect the prediction outcomes. This study is beneficial not only to industry players but also to investors and third-party advisors, as they can incorporate the prediction results into their decision-making processes.