Yunhai Tong

Correlating Twitter with the stock market through non-Gaussian SVAR

In this paper, we aim at studying the correlation between Twitter and the stock market. Specifically, we first apply non-Gaussian SVAR (structural vector autoregression) to identify possible relationships among the Twitter and stock market factors. Compared with conventional models such as Granger causality method which assume that the error items are Gaussian and only consider time-lag effect, non-Gaussian SVAR is under the assumption that the error items are non-Gaussian, better fitting the data in the stock market, and takes both instantaneous and time-lagged effects into account. We also visualize some distinctive relationships in parallel coordinates which is a well-developed multivariate visualization technique but seldom used in financial studies to the best of knowledge. Then, with the purpose of examining whether the Twitter-stock market relationship returned by non-Gaussian SVAR can help predict the stock market indicators, we build a series of regression models to predict DJI (Dow Jones Industrial Average Index) return in a sliding time window. Our experiments demonstrate that all the Twitter factors correlate with DJI return, and only the negative sentiment in tweets (posts on Twitter) is associated with DJI return volatility. Moreover, the lagged Twitter factors are more effective than the lagged stock market indicators in terms of predicting DJI return in the period of our data set.

Causal inference from financial factors: Continuous variable based local structure learning algorithm

For identifying the interrelationships of financial factors, we present a local structure learning based framework for Bayesian networks (BN) discovery from a large amount of continuous financial data without making parametric assumption. First, the skeleton of BN structure is learned by finding the parent and child set of each variable. Second, to direct the edges, the v-structures are learned by finding the spouse set of each node. To make the algorithm more useful to practitioners, our previously developed two-step accelerated method is incorporated into each step of local learning. Empirical studies on 56 US financial factors show both the efficiency and the effectiveness of our method.

NetRating: Credit Risk Evaluation for Loan Guarantee Chain in China

Guaranteed loans are a common way for enterprises to raise money from banks without any collateral in China. The enterprises are highly intertwined with each other, and hence form a densely connected guarantee network. As the economy is down in recent years, the default risk spreads along with the guarantee relations, and has caused great financial risk in many regions of China. Thus it puts forward a new challenge for financial regulators to monitor the enterprises involved in the guarantee network and control the system risk. However, the traditional financial risk management are based on vector space models, and could not handle the relations among enterprises. In this paper, based on the k-shell decomposition method, we propose a novel risk evaluation strategy, NetRating, to assess the risk level of each enterprise involved in the guaranteed loans. Besides, to deal with the direct guarantee networks, we propose the directed k-shell decomposition method, and extend NetRating strategy to the directed NetRating strategy. The application of our strategy in the real data verifies its effectiveness in credit assessment. It indicates that our strategy can provide a novel perspective for financial regulators to monitor the guarantee networks and control potential system risk.

A Structural Based Community Similarity Algorithm and Its Application in Scientific Event Detection

Graph similarity has been a crucial topic in network science, and is widely used in network dynamics, graph monitoring and anomalous event detection. However, few studies have paid attention to community similarity. The fact that communities do not necessarily own sub-modularity structure determines that graph similarity algorithms can not be applied to communities directly. Besides, the existing graph similarity algorithms ignore the organization structure of networks. Two communities can be regarded as the same when both their vertices and structure are identical. Thus the existing algorithms are unable to detect anomalous events about the shift of communities’ organization structure. In this paper, we propose a novel community similarity algorithm, which considers both the shift of vertices and the shift of communities’ layered structure. The layered structure of communities categorizes nodes into different groups, depending on their influence in the community. Both the influence of each node and the shift of nodes’ influence are expected to affect the similarity of two communities. Experiments on the synthetic data show that the novel algorithm performs better than the state-of-art algorithms. Besides, we apply the novel algorithm on the scientific data set, and identify meaningful anomalous events occurred in scientific mapping. The anomalous events are proved to correspond to the transition of topics for journal communities. It demonstrates that the novel algorithm is effective in detecting the anomalous events about the transition of communities’ structure.

Online adaptive method for disease prediction based on big data of clinical laboratory test

To better utilize the medical data in electronic medical records (EMR), this study aims to present an online adaptive method for disease prediction based on the medical data of clinical laboratory test (CLT) items stored in EMR. We firstly extract the diagnosis and CLT items information from the system, and then divide the CLT items into three categories to establish the patterns of CLT items, which are subsequently used for the selection of candidate diseases. A binary relevance approach based on logistic sparse group lasso method is finally used for disease prediction. Four groups of 21,288 patients with diagnosis of chronic hepatitis, hyperuricemia, hyperlipidemia and random diseases are used to test the performance of our method. Results show that the accuracy and recall for these four groups are all above 70%. As a primary attempt to practice intelligent healthcare, this model may have the potential values of computer-aided diagnosis. Further studies are suggested to combine CLT with other types of EMR data to further improve the prediction performance.

Adaptive logistic group Lasso method for predicting the no-reflow among the multiple types of high-dimensional variables with missing data

The prediction of no-reflow phenomenon aroused much attention, because of its independent association with increased in-hospital mortality, malignant arrhythmias, and cardiac failure. Many studies on prediction of no-reflow were carried out focusing on only few predictors. As big data era has been coming, high-dimensional predictors are available for prediction. However, as a common problem, big data analytics in healthcare from the electronic medical record (EMR) system is faced with many challenges, such as missing data processing, multiple types of variables processing and the high-dimensional data prediction. A general method based on improved weighted K-nearest neighbors and adaptive logistic group Lasso was proposed for predicting the no-reflow after cardiac surgery among the multiple types of variables with missing data. Compared with logistic regression, Lasso method, and artificial neural network method, our method has lower misclassification error rate and less complex model for no-reflow prediction, especially when predicting among multiple types of variables with missing data.

A novel dynamic community detection algorithm based on modularity optimization

Dynamic community detection has been an attractive topic due to its ability to reveal the evolutionary trends over time. However, existing dynamic community detection algorithms suffer from several disadvantages. Some make strong assumptions about the generation of communities, or require priori knowledge. In this paper, we propose a novel algorithm, dynamic Louvain method, to detect communities in temporal networks based on modularity optimization. The basic motivation is that the communities across different time steps should smoothly evolve. When partitioning temporal networks at a given time step, we should take historical network structure into consideration. Besides, this algorithm makes no assumption about the generation of communities, and is able to decide the number of communities automatically. This novel algorithm is applied to the temporal financial networks, and numerical evaluations show that this novel algorithm could obtain better partitions, compared with other state-of-art algorithms.

Identifying Key Drivers of Return Reversal with Dynamical Bayesian Factor Graph

In the stock market, return reversal occurs when investors sell overbought stocks and buy oversold stocks, reversing the stocks’ price trends. In this paper, we develop a new method to identify key drivers of return reversal by incorporating a comprehensive set of factors derived from different economic theories into one unified dynamical Bayesian factor graph. We then use the model to depict factor relationships and their dynamics, from which we make some interesting discoveries about the mechanism behind return reversals. Through extensive experiments on the US stock market, we conclude that among the various factors, the liquidity factors consistently emerge as key drivers of return reversal, which is in support of the theory of liquidity effect. Specifically, we find that stocks with high turnover rates or high Amihud illiquidity measures have a greater probability of experiencing return reversals. Apart from the consistent drivers, we find other drivers of return reversal that generally change from year to year, and they serve as important characteristics for evaluating the trends of stock returns. Besides, we also identify some seldom discussed yet enlightening inter-factor relationships, one of which shows that stocks in Finance and Insurance industry are more likely to have high Amihud illiquidity measures in comparison with those in other industries. These conclusions are robust for return reversals under different thresholds.

The history, hotspots, and trends of electrocardiogram

The electrocardiogram (ECG) has broad applications in clinical diagnosis and prognosis of cardiovascular disease. Many researchers have contributed to its progressive development. To commemorate those pioneers, and to better study and promote the use of ECG, we reviewed and present here a systematic introduction about the history, hotspots, and trends of ECG. In the historical part, information including the invention, improvement, and extensive applications of ECG, such as in long QT syndrome (LQTS), angina, and myocardial infarction (MI), are chronologically presented. New technologies and applications from the 1990s are also introduced. In the second part, we use the bibliometric analysis method to analyze the hotspots in the field of ECG-related research. By using total citations and year-specific total citations as our main criteria, four key hotspots in ECG-related research were identified from 11 articles, including atrial fibrillation, LQTS, angina and MI, and heart rate variability. Recent studies in those four areas are also reported. In the final part, we discuss the future trends concerning ECG-related research. The authors believe that improvement of the ECG instrumentation, big data mining for ECG, and the accuracy of diagnosis and application will be areas of continuous concern.

Efficient and effective Bayesian network local structure learning

In this paper, we propose a more efficient Bayesian network structure learning algorithm under the framework of score based local learning (SLL). Our algorithm significantly improves computational efficiency by restricting the neighbors of each variable to a small subset of candidates and storing necessary information to uncover the spouses, at the same time guaranteeing to find the optimal neighbor set in the same sense as SLL. The algorithm is theoretically sound in the sense that it is optimal in the limit of large sample size. Empirical results testify its improved speed without loss of quality in the learned structures.