Page - 37 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

tal adequacy level fromtheperspectiveofbanks.Thestate-of-art technologyofbigdata andartificial intelligencecanautomatically train interestingmodels fromalargeamount of structured and unstructured data, which breaks through the limitation of traditional predictingmethodsandgreatly improve thepredictionaccuracy. In this paper, we explore several machine learning methods on traditional struc- turedfinancialdataandtext-basedfinancialnews, topredict thecapitaladequacylevelof banks.Section2 introduces theworkflowand the relatedwork.Section3 introduces the predictionarchitectureandprocessingof the structuredandunstructureddata.Section4 explains theexperimentprocess andfivekindsofmachine learningalgorithmsexplored inour study.Section5analyzesandcompares theexperimental results. 2. RelatedWork Researches on the prediction of adequacy of capital and liquidity of commercial banks has been widely undertaken in recent decades. Diamond, Dybvig [1] (1983) believed that commercial banks were based on liquidity conversion, providing capital liquidity to themarket while facing a liquidity crisis caused by tight capital and liquidity gaps. The paper of Matz, Neu [2] (2007) put forward a pressure test model based on bank balance sheet, pointed out that banks should set the pressure scenario index of balance sheetbasedonhistoricaldata in thepressure test, andfinallyestimated theexpectedcash flow that banks can afford under different pressure scenarios. Since January 4th, 2007, China has officially operated the Shanghai InterbankOfferedRate (Shibor) to promote the rapid development of themoneymarket. Shibor is a barometer of the adequacy of bankcapital,withShiborupward representing the tight capitalmarket and the reverse is the loosecapitalmarket. With the rapid development of artificial intelligence technology, many researches began touseartificial intelligence technology tostudyfinancialproblems, especiallyus- ing text-baseddata.Asoneof theclassic scenes in thefieldofnatural languageprocess- ing (NLP), textcategorizationhasaccumulateda largenumberof technical implementa- tionmethods.The implementationapproachcanbe roughlydivided into twocategories: text classificationbasedon traditionalmachine learning and text classificationbasedon deep learning.R.Batra,S.M.Daudpota [3] (2018) focuson techniques involving senti- ment analysis inpredicting stock trends.FuliFenget al. [4]predict the stockmovement withanewmachine learningsolution.XiaoDingetal. [5] (2014)applied thedeeplearn- ingmethod to predict stock. They proposed a newmethod of event extraction, which extracted events from the news as input to the neural network. Sundermeyer et al. [6] (2012)used theLSTM(LongShort-TermMemory)unit to construct a languagemodel, anddiscovered itspotential in the languagemodel. 3. PredictionArchitecture This paper is based on both the traditional and deep learning text classification. Fig.1 shows the workflow of the prediction, which is composed of Data Processing,Model Training&Predicting andResult Evaluation&Analysis. TheData Processing is com- posedofDataPreprocessing, Feature&KeywordsExtraction andDataDimensionRe- Y.Duetal. /Predicting the InterbankCapitalAdequacyLevelBasedonFinancialDataAnalysis 37