北京雁栖湖应用数学研究院

The economy can be considered as a complex adaptive system in which the agents interact with each other in a complicated manner, making it incapable to deduce the behavior of the aggregate from the behavior of the average, or “representative” individual. In this talk, I will be presenting an emerging tool to reconstruct informative, dynamic, omnidirectional, and personalized networks (idopNetwork) from a population of economic data. idopNetwork provide a platform to explore systematically not only the internal complexity of a particular economy, leading to the identification of key modules and pathways, but also the mechanistic relationships among apparently distinct economic units. This tool can address how the economy or a market self-organizes, how this self-organization is affected by individual decisions in the economy and how the economy changes from state to state. Advances in network modeling are essential for disentangling economic complexity and functioning. Joint work with Yu Wang.

In recent years Internet of Things (IoT) devices are used for monitoring and controlling operations in urban and rural infrastructures including remote health monitoring. For remotely monitoring a patient only the health information at different time points is not sufficient, but the predicted values of the biomarkers (for future time points) are also required. In this article we propose a powerful statistical model for an efficient dynamic patient monitoring using wireless sensor nodes through Bayesian Learning (BL). A set of correlated biomarkers are measured from a patient over time where the sensors only report the ordinal outcomes (say; good, fair, high, very high) based on some prefixed thresholds. The challenge is to use these ordinal outcomes for monitoring and predicting the health status of the patient. We propose a linear mixed model where inter-biomarker correlations and intra-biomarker dependence are modeled simultaneously. The estimated and the predicted values of the biomarkers are transferred over internet so that the healthcare providers as well as the family members of the patient can remotely monitor the patient. Extensive simulation studies are performed to assess practical usefulness of our proposed joint model, and performance of the proposed joint model is compared to that of the other traditional models used in the literature.

Structural models in economics can offer appealing insights but often suffer from a poor fit with the data. In contrast, machine learning models offer rich flexibility but tend to suffer from over-fitting. We propose a novel framework that incorporates useful economic restrictions from a structural model into a machine learning model through transfer learning. The core idea is to first construct a neural-network representation of the structural model, and then update the network using information from real data. In an example application to option pricing, our hybrid model significantly outperforms both the structural model and a conventional deep-learning model. The out-performance of the hybrid model is more significant when the sample size of real data is limited or under volatile market conditions.

For misspecified models, from predictive viewpoint, generally, three are three different predictive distributions for candidate use, i.e., plug-in predictive distribution, the regular Bayesian predictive distribution, and the sandwich Bayesian posterior predictive distribution where the regular posterior distribution is substituted by sandwich posterior distribution proposed by Mu¨ller(2013) (Econometrica, 81(5), 1805-1849). First, on the basis of Kullback-Leibler (KL) loss function, we show that the sandwich Bayesian predictive distribution can yield lower asymptotic risk than the regular Bayesian predictive distribution. Those results extend the conclusion by Mu¨ller(2013) on the parameter estimation into the predictive distributions. Furthermore, we give the conditions that the sandwich Bayesian posterior predictive distribution is better or not than the plug-in predictive distribution. Second, under frequent risk analysis, based on these two Bayesian predictive distributions, we proposed some important information criterion for comparing misspecified models which can be unbiased estimators for the risks based on these two predictive distributions. Third, we established the relationship between the propose information criterion and the existing information criterion such as the popular AIC, TIC, and DICs, etc. At last, we illustrate the proposed new information criterion using some real studies in economics and finance.

In this paper, we propose a new method – subsampling-based random projection (SRP) regression – for analyzing datasets comprising a very large number of observations. The main advantage of this approach is that it only uses the sum information of each sub-sample, which greatly reduces the memory requirements of our procedure, enabling computation even on modest-sized laptop computers. We develop large sample theory for these estimators, and simulations demonstrate the good finite sample performance of our method. We also provide an application to a large dataset (comprising billions of observations) drawn from a mobile advertising platform.

We apply economic narratives to macroeconomic forecasting using a large news corpus and machine learning algorithms. We measure economic narratives quantitatively from the full text content of Wall Street Journal articles and articles represent them as interpretable news topics. The results indicate that narrative-based forecasts are more accurate than the benchmarks both in-sample and out-of-sample. Overall, we highlight the important role of economic narratives in economic forecasting.

The mining industry is one of the most dangerous industries worldwide. Typical accidents include but are not limited to gas explosions, floods, and derailing. Unfortunately, most of these accidents are associated with human errors, often due to miners’ unsafe acts. Thus, timely monitoring and correcting miners’ unsafe acts is of great importance for safety assurance. To this end, we develop here a double smoothing kernel estimation method. It takes high-resolution images as inputs and then detects miners in the images automatically. Compared with the classical kernel density estimator, the new method contains two layers of nonparametric kernel smoothing. We show theoretically that the resulting density estimator enjoys a much improved statistical efficiency, but also suffers from high computational cost. To speed up the computation, a grid point approximation (GPA) method is further developed. Once a miner is detected in the image, a pre-trained MobileNet model (i.e., a classical deep learning model based on convolutional neural networks) can be used to extract feature vectors. Based on the extracted feature vectors, a standard logistic regression model can be trained to classify the miners’ acts into safe or unsafe categories. The resulting out-of-sample prediction accuracy is excellent.

This talk proposes a consumption-based model to explain puzzling unstable, i.e., sometimes positive and sometimes negative, relations between stock market variance with both market risk premia and prices. In the model, market risk premia depend positively (negatively) on “fear” (“euphoria”) variance. Market prices, which decrease with discount rates, correlate negatively (positively) with fear (euphoria) variance. Because it is the sum of fear and euphoria variances, market variance may correlate positively or negatively with expected returns and prices, depending on the relative importance of the two variances. Our empirical results support model’s key assumptions and many novel implications.

This talk will present a new perspective on revealing the selection, train and test approaches of deep learning methods in biomedical applications. First is a deep learning approach combining graph representation and neural networks for multi-mer protein interaction prediction. Second is using implicit neural networks to model the continuous time and space changes of protein dynamics. Third is a pre-trained big model for multiple purpose protein computations, which has been validated by experimental results. Our works will help to understand and develop artificial intelligence algorithms for diverse applications.

We study the contribution of Chinese science to US technology, as evidenced by patent citation of scientific publications. Our analysis documents that foreign science is prevalent in US patents, with China emerging as a leading foreign producer of science cited in US patents. In 2020, nearly 30% of US science-reliant patents referenced articles published by Chinese scholars, representing a five-fold increase compared to 2000. While Chinese science has predominantly played a supplementary role in US technological advancement, we observe that in nearly 15% of US patents citing Chinese science, China serves as the primary source of referenced scientific knowledge. Furthermore, the influence of Chinese science is not limited to US innovators, as approximately 14% of USPTO-granted foreign, science-reliant patents have referenced Chinese science.

Bridging neural networks with partial differential equations holds significant importance, as it not only enhances the interpretability of neural networks but also sheds light on designing network architectures. In this talk, we establish convection-diffusion equation models based on rigorous theoretical analysis. The convection-diffusion equation model not only covers existing network structures, but also illuminates novel network design, COnvection dIffusion Networks (COIN). Numerical results demonstrate the effectiveness of COIN in various benchmarks, as well as its potential in novel tasks such as disease prediction.

This study constructs a Dynamic Stochastic General Equilibrium (DSGE) model, incorporating a CBDC featuring managed anonymity and interest-bearing rules, to study the influence of CBDC issuance on economic fluctuations. We find that the introduction of CBDC, by affecting the automatic stabilizer function of taxation, can modulate short-term economic fluctuations. However, the effects hinge on the interest-bearing rules and anonymity degree of the CBDC. Specifically, introducing a managed anonymous CBDC, with interest rates pegged to policy rates, increases economic fluctuations caused by exogenous shocks. Thus, contrary to prevailing literature, the Taylor-like interest rule may not be optimal when considering managed anonymity. Furthermore, our findings reveal a non-linear, inverted U-shaped relationship between CBDC anonymity and the enhancement of social welfare. Based on these insights, we propose principles for optimizing CBDC design, and our quantitative analyses suggest that introducing such an optimized CBDC better regulates short-term economic fluctuations resulting from various exogenous shocks, thereby improving social welfare.

In the problem of joint pricing and inventory management the retailer makes simultaneously a price decision and an inventory order-up-to decision at the beginning of each review period. The demands are being modeled as either a parametric or nonparametric function depending on the prices. <br>In this talk, I will introduce two of my recent works advancing this problem: the first one deals with fixed ordering costs under the backlogging setting, with a parametric (generalized linear) demand model. The second one studies nonparametric demand models with censored demands and lost sales. The techniques involved include a novel UCB analysis over trajectories of (s,S,p) policies, and a noisy comparison oracle constructed for censored demand models.<br>This talk is based on the following two papers:<br>[1] Chen, Boxiao, David Simchi-Levi, Yining Wang, and Yuan Zhou. “Dynamic pricing and inventory control with fixed ordering cost and incomplete demand information.” Management Science 68, no. 8 (2022): 5684-5703.<br>[2] Chen, Boxiao, Yining Wang, and Yuan Zhou. “Optimal policies for dynamic pricing and inventory control with nonparametric censored demands.” Management Science 70, no. 5 (2024): 3362-3380.

AI in Finance is a fascinating cross-disciplinary area. However, a significant gap exists between these two domains, characterized by distinct cultures and objectives. In this presentation, our goal is to bridge this gap by highlighting the unique aspects of machine learning in finance and introducing novel methodologies with successful applications. We will cover three key topics: uncertainty quantification, generative learning, and model stability in finance. We contend that end-to-end black-box models are generally ineffective in this context. Instead, models elaborately designed to capture the natures of financial data, such as uncertainty, dependency structures, and tail properties, are more likely to succeed. The methodologies we have proposed achieve state-of-the-art performance in tasks such as risk forecasting and portfolio construction.

Knowledge graphs (KGs) consist of many facts that connect real-world entities with various relations. Due to the prevalence of relational data in practice, KG has a wide range of applications including recommendation systems, natural language processing (NLP) and question answering (QA), etc. In this talk, I will first present MQuadE, a embedding-based method for completing the KG. Experiments show that MQuadE outperforms the SOTA of embedding based methods. Then, I will introduce MLN4KB, an efficient Markov logic network (MLN) engine for large scale knowledge graphs. Experiments demonstrate that MLN4KB is orders of magnitudes faster than existing MLN engines. Lastly, I will combine embedding based with a relaxed version of MLN, which comes up with a neuro-symbolic reasoning engine called DiffLogic. On benchmark datasets, we empirically show that DiffLogic surpasses baselines in both effectiveness and efficiency.