財務報表舞弊對於投資大眾影響甚鉅，而投資大眾之決策依據大多仰賴企業所提供之財務報表，若管理階層寓意提供不正確或偏誤財務資訊，對投資大眾的傷害是相當深遠，對整個金融市場影響更是不可言喻，有鑑於會計審計人員為財務報表最後防線，且其若出具不正確查核報告，即審計失敗，會面臨相當龐大之訴訟風險，而大多會計審計人員缺乏相關經驗與知識，導致出具不正確/不適當查核報告，再加上決策者隸屬於資訊爆炸之大數據時代，更多資訊充斥在其決策環境中，亦會導致其決策判斷之可能性，為了降低此風險以及穩定金融市場，本計畫欲提出一個以人工智慧技術為基礎之動態財務報表舞弊偵測架構，此架構利用投資組合概念對財務數據做波動性調整，以降低異常波動對數據之影響，年報中之文字資訊對會計準則選取有相當多之論述，將以文字探勘技術萃取此資訊，而社群媒體已是決策者獲取資訊的主要來源，此架構將採用社群媒體探勘萃取相關的主題資訊，以降低資訊超載之困境，並進一步採用動態網絡型態資料包絡分析法萃取經理人之管理能力，並彙整所有資訊於人工智慧技術以建構財務報表舞弊偵測模型，最後利用佔約約略集合萃取其內部決策邏輯，以利決策者做更精闢之判斷與分析。
Fraudulent financial statement (FFS) detections are essential and critical in order to protect stock market participants as well as sound the stability of economic market. In recent years, FFS have begun to appear and continue to grow dramatically and has ruined the confidence of market participants and slowed down the development of the global markets. Auditors are the final goal-keepers of financial statements who have to yield audited financial statements. Most proportion of investors relied on audited financial statements to form their final decisions. If they provided inappropriate or wrong audited reports, they will encounter tremendous uncertainty and risk. However, many auditors lack the experience and expertise to deal with the related risks. Thus, this study introduces an artificial intelligence (AI)-based dynamic FFS detection architecture to ameliorate these risks. The architecture integrates volatility-adjusted strategy, text mining (TM), social media mining (SMM), dynamic network data envelopment analysis (DNDEA), fuzzy K means (FKM), support vector machine (SVM), dominance-based rough set theory (DRST) and multiple criteria decision makings (MCDM). The volatility-adjusted strategy is used to ameliorate the impact of great fluctuation in financial ratios. The TM and SMM are used to extract the relevant and real-time textual information for decision making procedures. The DNDEA model is applied to extract the managers’ managerial ability. The FKM is performed to select the essential instances and then fed into SVM to construct the FFS detection model. However, the SVM-based detection model with superior forecasting performance comes with a critical challenge of lacking interpretability. If the forecasting model has a bad interpretability for users, it will impede its real-life applications. To overcome this problem, the DRST is executed to extract the inherent decision logics from SVM-based model. The MCDM is used to determine which knowledge extraction method can reach the optimal performance in a logical and systematic way.
