Machine Learning and Predictive Analytics in Accounting, Finance and Management
Permanent URI for this collection
Browse
Recent Submissions
1 - 7 of 7
-
ItemTowards Optimal Free Trade Agreement Utilization through Deep Learning Techniques( 2020-01-07)In recent years, deep learning based methods achieved new state of the art in various domains such as image recognition, speech recognition and natural language processing. However, in the context of tax and customs, the amount of existing applications of artificial intelligence and more specifically deep learning is limited. In this paper, we investigate the potentials of deep learning techniques to improve the Free Trade Agreement (FTA) utilization of trade transactions. We show that supervised learning models can be trained to decide on the basis of transaction characteristics such as import country, export country, product type, etc. whether FTA can be utilized. We apply a specific architecture with multiple embeddings to efficiently capture the dynamics of tabular data. The experiments were evaluated on real-world data generated by Enterprise Resource Planning (ERP) systems of an international chemical and consumer goods company.
-
ItemEnhancing Customer Satisfaction Analysis with a Machine Learning Approach: From a Perspective of Matching Customer Comment and Agent Note( 2020-01-07)With the booming of UGCs, customer comments are widely utilized in analyzing customer satisfaction. However, due to the characteristics of emotional expression, ambiguous semantics and short text, sentiment analysis with customer comments is easily biased and risky. This paper introduces another important UGC, i.e., agent notes, which not only effectively complements customer comment, but delivers professional details, which may enhance customer satisfaction analysis. Moreover, detecting the mismatch on aspects between these two UGCs may further help gain in-depth customer insights. This paper proposes a machine learning based matching analysis approach, namely CAMP, by which not only the semantics and sentiment in customer comments and agent notes can be sufficiently and comprehensively investigated, but the granular and fine-grained aspects could be detected. The CAMP approach can provide practical guidance for following-up service, and the automation can help speed-up service response, which essentially improves customer satisfaction and retains customer loyalty.
-
ItemAutomating Lead Scoring with Machine Learning: An Experimental Study( 2020-01-07)Companies often gather a tremendous amount of data, such as browsing behavior, email activities and other contact data. This data can be the source of important competitive advantage by utilizing it in estimating a contact's purchase probability using predictive analytics. The calculated purchase probability can then be used by companies to solve different business problems, such as optimizing their sales processes. The purpose of this article is to study how machine learning can be used to perform lead scoring as a special application case of making use of purchase probability. Historical behavioral data is used as training data for the classification algorithm, and purchase moments are used to limit the behavioral data for the contacts that have purchased a product in the past. Different ways of aggregating time-series data are tested to ensure that limiting the activities for buyers does not result in model bias. The results suggest that it is possible to estimate the purchase probability of leads using supervised learning algorithms, such as random forest, and that it is possible to obtain business insights from the results using visual analytics
-
ItemA CNN Based System for Predicting the Implied Volatility and Option Prices( 2020-01-07)The evaluations of option prices and implied volatility are critical for option risk management and trading. Common strategies in existing studies relied on the parametric models. However, these models are based on several idealistic assumptions. In addition, previous research of option pricing mainly depends on the historical transaction records without considering the performance of other concurrent options. To address these challenges, we proposed a convolutional neural network (CNN) based system for predicting the implied volatility and the option prices. Specifically, the customized non-parametric learning approach is first used to estimate the implied volatility. Second, several traditional parametric models are also implemented to estimate these prices as well. The convolutional neural network is utilized to obtain the predictions based on the estimation of the implied volatility. Our experiments based on Chinese SSE 50ETF options demonstrate that the proposed framework outperforms the traditional methods with at least 40.12% performance enhancement in terms of RMSE.
-
ItemAn Explicative and Predictive Study of Employee Attrition Using Tree-based Models( 2020-01-07)We develop tree-based models to estimate the probability of an employee leaving a firm during a job transition from a dataset of anonymously submitted resumes through Glassdoor’s online portal. Dataset construction and summary statistics are first summarized followed by a more in depth examination through four exploratory studies. Insights provided by these studies are then used to engineer features that serve as input into subsequent attrition related predictive models. We finally perform a thorough search through several dozen binary classification techniques in the cases of an original and extended feature set. We find tree-based methods including random forests and light gradient boosted trees provide the overall strongest predictive performance. Finally, we summarize ROC curves for several such models and describe future potential research directions.
-
ItemImproving Credit Risk Analysis with Cluster Based Modeling and Threshold Selection( 2020-01-07)Credit risk has been an integral part of financial industry and is a challenging and difficult risk to manage. The diverse behavior of borrowers adds challenges to the risk analysis. Failing to accurately identify the borrowers' risk can lead to huge investment losses. Credit scoring is a popular and commonly used technique to analyze credit risk. A single credit scoring model may not be capable of generating a common rule to classify borrowers and hence segmented modeling can be applied to create more specific classification rules for achieving higher classification accuracy. In this study segmented modeling is applied with threshold selection for each segment to reduce relative cost of misclassification. The results from the study show that threshold selection based on the segmented modeling can give improvement over a single credit scoring model.
-
Item