以往國內有關退休金給付之研究，大多著重於退休金制度的實施方式、退休金提撥方法、及退休基金運用等，但對於事業單位實際提存行為之管理缺乏相關研究。本研究主要目的在利用勞工退休準備基金戶相關資料，探討基金戸之特徵，並建構勞工退休準備基金戸流失預測模式，以協助主管機關落實勞工退休準備金之提撥、監督、與管理。本研究利用勞工退休準備金基金戶存儲、給付等相關資料，進行選擇、編碼、正規化及縮減等前置處理，再應用資料探勘技術，探討基金戶隱藏之存儲模式、給付模式，並建構出流失預測模式，以便在基金戶存儲、給付行為發生異常時，即發出警訊，強化對基金戶稽催管理，以保障勞工權益。本研究之資料為2003年9月之基金戶共31517筆，經前置處理篩選後之資料共12772筆。先利用叢集法進行基金戶特徵及行為模式分析，再分別以決策樹、及類神經網路建構基金戶流失預測模式。研究結果發現：1. 整體儲存次數及金額皆有偏低現象，顯示整體基金戶儲備能力不佳。2. 整體基金戶之支應偏高，顯示企業所提繳之數額，無法支應勞工退休金需要。3. 本研究建構之勞工退休準備金流失預測模式主要影響因素為繳款評分、最近儲存日、開戶年限、成立年限及儲存次數。其中易成為預警戶的指標有成立年限大於13年、或開戶年限大於8年、或繳款評分低於58分、或最近儲存日為4-12個月、或久未儲存卻有給付行為等。4. 本研究同時發現利用叢集分析結合決策樹或類神經網路，可獲得較佳之分類模式。

Most of the related researches about retired pension system are focused on the implementation of retirement system. There are short of the researches about how to monitor and manage the worker pension account. The goal of this research is to use the data of worker pension account to analysis the account characteristics and construct the worker pension account churn prediction model. The samples are worker pension account records of September 2003. First, we do the data cleaning, and get 12772 records out from 31517 records. Then, we apply two stages clustering analysis to analyze the enterprise’s demographic characteristics and behavior. Finally, we utilize the decision tree and neural network techniques to construct the prediction model. In this research, we got some conclusions:
1. The capability of whole reservation was insufficient because that both the deposit frequency and amount were pretty low. 2. The results showed that the worker pension payment is too much. It indicated that the amount of submission can not afford the worker pension payment. 3. The results showed that the critical factors of prediction churn were lifetime of enterprise, lifetime of fund account, and deposit score. If the workers pension account with lifetime of enterprise is greater than 13 years, or lifetime of fund account is greater than 8 years, or the deposit score is less than 58 points, or F value equals 2 or 3, tends to be churn. 4. We also found that associated the clustering method with decision tree or neural network method can get a better prediction model.