PUBlications

Figure: The framework of RFNs for Asyn-MTS. It has three component variables X^1_t , X^2_t , X^3_t . The solid dots are observations. Different colors refer to different variables. In the marginal block, each variable has its own set of hidden states h^1(t), h^2(t), h^3(t). The hidden states of a particular variable will be updated only when it has an observation. In the multivariate block, the time-tk base distribution parameters, µ_{t_k} and Σ_{t_k} , are functions of the hidden states h_{t_{k−}} = [h^_{t_{k−}}; h^2_{t_{k−}}; h^3_{t_{k−}}] and additionally the component variables X^{−d}_{t_k} := [x^1_{t_k}, · · · , x^{d−1}_{t_k}]. 

Estimating treatment effects from observational data is subject to a covariate shift problem incurred by selection bias. Recent research has sought to mitigate this problem by balancing the distribution of representations between the treated and controlled groups. The rationale behind this is that counterfactual estimation relies on (1) preserving the predictive power of factual outcomes and (2) learning balanced representations. However, there is a trade-off between achieving these two objectives. In this paper, we propose a novel model, DIGNet, which is designed to capture the patterns that contribute to outcome prediction (task 1) and representation balancing (task 2) respectively. Specifically, we derive a theoretical upper bound that links the concept of propensity confusion to representation balancing, and further transform the balancing Patterns into Decompositions of Individual propensity confusion and Group distance minimization (PDIG) to capture more effective balancing patterns. Moreover, we suggest decomposing proxy features into Patterns of Pre-balancing and Balancing Representations (PPBR) to preserve patterns that are beneficial for outcome modeling. Extensive experiments confirm that PDIG and PPBR follow different pathways to achieve the same goal of improving treatment effect estimation. We hope our findings can be heuristics for investigating factors influencing the generalization of representation balancing models in counterfactual estimation. 

Causal learning is the key to obtaining stable predictions and answering what if problems in decision-makings. In causal learning, it is central to seek methods to estimate the average treatment effect (ATE) from observational data. The Double/Debiased Machine Learning (DML) is one of the prevalent methods to estimate ATE. However, the DML estimators can suffer from an error-compounding issue and even give extreme estimates when the propensity scores are close to 0 or 1. Previous studies have overcome this issue through some empirical tricks such as propensity score trimming, yet none of the existing works solves it from a theoretical standpoint. In this paper, we propose a Robust Causal Learning (RCL) method to offset the deficiencies of DML estimators. Theoretically, the RCL estimators i) satisfy the (higher-order) orthogonal condition and are as consistent and doubly robust as the DML estimators, and ii) get rid of the error-compounding issue. Empirically, the comprehensive experiments show that: i) the RCL estimators give more stable estimations of the causal parameters than DML; ii) the RCL estimators outperform traditional estimators and their variants when applying different machine learning models on both simulation and benchmark datasets, and a mimic consumer credit dataset generated by WGAN. 

Unsupervised domain adaptation enables knowledge transfer from a labeled source domain to an unlabeled target domain by aligning the learnt features of both domains. The idea is theoretically supported by the generalization bound analysis in Ben-David et al. (2007), which specifies the applicable task (binary classification) and designates a specific distribution divergence measure. Although most distribution-aligning domain adaptation models seek theoretical grounds from this particular bound analysis, they do not actually fit into the stringent conditions. In this paper, we bridge the long-standing theoretical gap in literature by providing a unified generalization bound. Our analysis can well accommodate the classification/regression tasks and most commonly-used divergence measures, and more importantly, it can theoretically recover a large amount of previous models. In addition, we identify the key difference in the distribution divergence measures underlying the diverse models and commit a comprehensive in-depth comparison of the commonly-used divergence measures. Based on the unified generalization bound, we propose new domain adaptation models that achieve transferability through domain-invariant representations and conduct experiments on real-world datasets that corroborate our theoretical findings. We believe these insights are helpful in guiding the future design of distribution-aligning domain adaptation algorithms. 

Figure: Visualization of three 2-D Gaussian distributions. The KL divergences between P^X and P^Y and between P^Z and P^Y are the same, which is 1/2. However, P^X differs from P^Y in the 2nd marginal distribution, whereas the difference between P^Z and P^Y lies in the covariance matrix. This example illustrates that a single divergence measure based the joint distribution alone cannot distinguish whether the distributional differences come from the marginals or the dependence structure. 

We propose a counter-cyclical initial margin model for option portfolios. Our model explores the intrinsic netting within a given portfolio of European options and outputs a constant upper bound of the maximum possible loss. This feature would allow option clearinghouses and regulators to gauge the tightest margin levels that are stable. We compare our model with the scenario-based SPAN model and the sensitivity-based SIMM model in terms of the netting efficiency and the procyclical property. Using the SPX options and the interest rate swaptions as examples, we quantify the minimum amount of additional margins needed to make them fully counter-cyclical. We then show how to strike a balance between risk-sensitivity and counter-cyclicality if needed by mixing our model flexibly with a prevailing risk-sensitive margin model. 

This paper develops a conditional quantile model that can learn long term and short term memories of sequential data. It builds on sequential neural networks, but can outputs interpretable dynamics, and consistenly outperforms the GARCH family as well as models using filtered historical simulation, conditional extreme value theory, and dynamic quantile regression for Value-at-Risk forecasts. Upon applying the model to asset return time series across eleven asset classes using historical data from the 1960s to 2018, it is revealed and confirmed that conditional quantiles of asset return have persistent sources of risk that are not coming from those responsible for volatility clustering. 

Subset selection identifies a subgroup of items from a given collection to achieve specific goals. Researchers typically choose between the regularization method and the perturbation method to make the selection operator differentiable in order to implement it in end-to-end learning frameworks. This paper unifies these two schemes through a probabilistic interpretation for regularization relaxation. We build some concrete examples to show the generic connection between these two relaxation, and we evaluate the perturbed selector as well as the regularized selector on two tasks: the maximum entropy sampling problem and the feature selection problem.

Double/Debiased Machine Learning (DML) method aims to remove the covariates' impact to obtain unbiased estimation of treatment effects. The essence of DML lies in the use of two possibly mis-specified models: one for predicting the treatment variable based on covariates and another for predicting the outcome variable based on covariates. The problem is existing DML estimators is prone to the error-compounding issue when the propensity scores are close to zero or one. This paper proposes a new estimator that is just as consistent and doubly robust as existing DML estimators but not susceptible to this problem. 

Surge pricing finds equilibrium prices in periods of excessive demand or scarce supply. Effective or not, it is a business risk when riders perceive spiking prices as exploitation of people's emergency. This paper studies subsidy policies that avoid the downside of surge pricing when accommodating the fluctuation of demand. We show that the re-usability of driver supply presents a hidden capacity. Tapping it wisely through supply-side subsidies prescribes a non-pricing alternative to the current pricing policies without the need for either hiking price or recruiting new drivers. We use a queueing model together with the Stackelberg game to analyze how to optimally subsidize a reusable pool of driver supply, both myopically and in the long run. Knowing drivers are self-interested and given a specific structure of base trip fare, our analysis shows that injecting a healthy dose of myopic subsidy into the matching process reverts unfavorable decisions of individual drivers. In aggregate, the induced supply multiplier effect significantly boosts vehicle circulation and effective demand. We also show that, on the other hand, the impact of myopic subsidies on long-run throughput is not monotonic. There is a physical limit in terms of how much the platform can tap this hidden capacity. However large the budget of incentive, its size is intrinsically constrained by the spatial structure of base trip fare and the distribution of customer travel distances.  

Invariant learning methods try to find an invariant predictor across several environments and have become popular in OOD generalization. However, in situations where environments do not naturally exist in the data, they have to be decided by practitioners manually. Environment partitioning, which splits the whole training dataset into environments by algorithms, will significantly influence the performance of invariant learning and has been left undiscussed. A good environment partitioning method can bring invariant learning to applications with more general settings and improve its performance. We propose to split the dataset into several environments by finding low-correlated data subsets. Theoretical interpretations and algorithm details are both introduced in the paper. Through experiments on both synthetic and real data, we show that our Decorr method can achieve outstanding performance, while some other partitioning methods may lead to bad, even below-ERM results using the same training scheme of IRM. 

We develop explicit asymptotic expansions of the portfolio Value-at-Risk (VaR) and portfolio Expected Shortfall (ES) for a large family of multivariate elliptical distributions. The family includes distributions of exponential type such as Kotz distributions, and power type such as the multivariate Student t-distribution. Our results imply that the difference between the portfolio ES and its VaR depends on the tail heaviness of the joint asset return distribution. For assets exhibiting exponential tail decay, the ratio between ES and VaR is asymptotically zero, whereas for assets exhibiting power type tail decay, the portfolio ES is strictly larger than its VaR. The amount of the risk reduction through merging subportfolios into a single portfolio depends solely on the dispersion of the joint asset return distribution. 

We present a stochastic-volatility, short rate term structure model, which extends the classic multi-factor Hull-White model. This model is designed to fit the swaption implied volatility cube and to incorporate the two-curve modeling paradigm. The model exhibits non-Gaussian forward swap rates whose distributions are parameterized across the dimensions of the volatility cube: underlying tenor, option strike and option expiration. To facilitate rapid model calibration, we establish suitable asymptotic expressions for the bond prices. Furthermore, we derive an effective SABR dynamics for each forward swap rate. Finally, we use the mean field approximation to match the effective SABR parameters corresponding to each swaption to the market levels. 

Traditional sentiment construction in finance relies heavily on the dictionary-based approach, with a few exceptions using simple machine learning techniques such as Naive Bayes classifier. While the current literature has not yet invoked the rapid advancement in the natural language processing, we construct in this research a textual-based sentiment index using a well-known pre-trained model BERT developed by Google, especially for three actively trading individual stocks in Hong Kong market with at the same time the hot discussion on Weibo.com. On the one hand, we demonstrate a significant enhancement of applying BERT in financial sentiment analysis when compared with the existing models. On the other hand, by combining with the other two commonly-used methods when it comes to building the sentiment index in the financial literature, i.e., the option-implied and the market-implied approaches, we propose a more general and comprehensive framework for the financial sentiment analysis, and further provide convincing outcomes for the predictability of individual stock return by combining LSTM (with a feature of a nonlinear mapping). It is significantly distinct with the dominating econometric methods in sentiment influence analysis which are all of a nature of linear regression.