Federated Learning enables a population of clients, working with a trusted server, to collaboratively learn a shared machine learning model while keeping each client's data within its own local systems. This reduces the risk of exposing sensitive data, but it is still possible to reverse engineer information about a client's private data set from communicated model parameters. Most federated learning systems therefore use differential privacy to introduce noise to the parameters. This adds uncertainty to any attempt to reveal private client data, but also reduces the accuracy of the shared model, limiting the useful scale of privacy-preserving noise. A system can further reduce the coordinating server's ability to recover private client information, without additional accuracy loss, by also including secure multiparty computation. An approach combining both techniques is especially relevant to financial firms as it allows new possibilities for collaborative learning without exposing sensitive client data. This could produce more accurate models for important tasks like optimal trade execution, credit origination, or fraud detection. The key contributions of this paper are: We present a privacy-preserving federated learning protocol to a non-specialist audience, demonstrate it using logistic regression on a real-world credit card fraud data set, and evaluate it using an open-source simulation platform which we have adapted for the development of federated learning systems.

In the financial sector, a reliable forecast the future financial performance of a company is of great importance for investors' investment decisions. In this paper we compare long-term short-term memory (LSTM) networks to temporal convolution network (TCNs) in the prediction of future earnings per share (EPS). The experimental analysis is based on quarterly financial reporting data and daily stock market returns. For a broad sample of US firms, we find that both LSTMs outperform the naive persistent model with up to 30.0% more accurate predictions, while TCNs achieve and an improvement of 30.8%. Both types of networks are at least as accurate as analysts and exceed them by up to 12.2% (LSTM) and 13.2% (TCN).

In this paper we propose an extension of the Merton model. We apply the subdiffusive mechanism to analyze equity warrant in a fractional Brownian motion environment, when the short rate follows the subdiffusive fractional Black-Scholes model. We obtain the pricing formula for zero-coupon bond in the introduced model and derive the partial differential equation with appropriate boundary conditions for the valuation of equity warrant. Finally, the pricing formula for equity warrant is provided under subdiffusive fractional Brownian motion model of the short rate.

We study the identification of channels of policy reforms with multiple treatments and different types of selection for each treatment. We disentangle reform effects into policy effects, selection effects, and time effects under the assumption of conditional independence, common trends, and an additional exclusion restriction on the non-treated. Furthermore, we show the identification of direct- and indirect policy effects after imposing additional sequential conditional independence assumptions on mediating variables. We illustrate the approach using the German reform of the allocation system of vocational training for unemployed persons. The reform changed the allocation of training from a mandatory system to a voluntary voucher system. Simultaneously, the selection criteria for participants changed, and the reform altered the composition of course types. We consider the course composition as a mediator of the policy reform. We show that the empirical evidence from previous studies reverses when considering the course composition. This has important implications for policy conclusions.

We develop a duality theory for the problem of maximising expected lifetime utility from inter-temporal wealth over an infinite horizon, under the minimal no-arbitrage assumption of No Unbounded Profit with Bounded Risk (NUPBR). We use only deflators, with no arguments involving equivalent martingale measures, so do not require the stronger condition of No Free Lunch with Vanishing Risk (NFLVR). Our formalism also works without alteration for the finite horizon version of the problem. As well as extending work of Bouchard and Pham to any horizon and to a weaker no-arbitrage setting, we obtain a stronger duality statement, because we do not assume by definition that the dual domain is the polar set of the primal space. Instead, we adopt a method akin to that used for inter-temporal consumption problems, developing a supermartingale property of the deflated wealth and its path that yields an infinite horizon budget constraint and serves to define the correct dual variables. The structure of our dual space allows us to show that it is convex, without forcing this property by assumption. We proceed to enlarge the primal and dual domains to confer solidity to them, and use supermartingale convergence results which exploit Fatou convergence, to establish that the enlarged dual domain is the bipolar of the original dual space. The resulting duality theorem shows that all the classical tenets of convex duality hold. Moreover, at the optimum, the deflated wealth process is a potential converging to zero. We work out examples, including a case with a stock whose market price of risk is a three-dimensional Bessel process, so satisfying NUPBR but not NFLVR.

We propose a new model for the joint evolution of the European inflation rate, the European Central Bank official interest rate and the short-term interest rate, in a stochastic, continuous time setting.

We derive the valuation equation for a contingent claim and show that it has a unique solution. The contingent claim payoff may depend on all three economic factors of the model and the discount factor is allowed to include inflation.

Taking as a benchmark the model of Ho, H.W., Huang, H.H. and Yildirim, Y., Affine model of inflation-indexed derivatives and inflation risk premium, (European Journal of Operational Researc, 2014), we show that our model performs better on market data from 2008 to 2015.

Our model is not an affine model. Although in some special cases the solution of the valuation equation might admit a closed form, in general it has to be solved numerically. This can be done efficiently by the algorithm that we provide. Our model uses many fewer parameters than the benchmark model, which partly compensates the higher complexity of the numerical procedure and also suggests that our model describes the behaviour of the economic factors more closely.

The lack of interpretability and transparency are preventing economists from using advanced tools like neural networks in their empirical work. In this paper, we propose a new class of interpretable neural network models that can achieve both high prediction accuracy and interpretability in regression problems with time series cross-sectional data. Our model can essentially be written as a simple function of a limited number of interpretable features. In particular, we incorporate a class of interpretable functions named persistent change filters as part of the neural network. We apply this model to predicting individual's monthly employment status using high-dimensional administrative data in China. We achieve an accuracy of 94.5% on the out-of-sample test set, which is comparable to the most accurate conventional machine learning methods. Furthermore, the interpretability of the model allows us to understand the mechanism that underlies the ability for predicting employment status using administrative data: an individual's employment status is closely related to whether she pays different types of insurances. Our work is a useful step towards overcoming the "black box" problem of neural networks, and provide a promising new tool for economists to study administrative and proprietary big data.

Pandemics such as COVID-19 have lethal potential for inflicting long-lasting cyclic devastations if required preventive, curative and reformative steps are not taken up in time which puts forth mammoth multi-dimensional challenges for survival before mankind. Scientists and policymakers all around are striving to achieve R $\leq$ 1 alongside having less number of CoVID-19 patients. Lockdowns all across the globe have been implemented for the sake of social physical distancing. However, even if the desired R value status is achieved it becomes nowhere near safe. As normal social activity and inter-regional travel resumes, danger of contraction of the virus from undetected asymptomatic carriers and reactivation of the virus in previously affected patients looms over. The virus poses further threat due to its chances of resurgence, its mutative and adaptive nature thereby giving limited medical respite. The problems intensify with increasing population density whilst varying with several socio-economic-geo-cultural and human activity parameters. Such zoonotic pandemics unravel the primary challenges of all countries in securing the general wellbeing of the society. Ensuring a mechanism for policy designs envisaging crisis scenarios through continuous analysis of real-time region-specific data about societal activities and disease/health indicators can be the only solution. An approach perspective is discussed for addressing the tightly-coupled UN Sustainable goals (2, 3, 6, 12 and 13) for developing a general-scale computational agent-based model to estimate the downstream and upstream measures for achieving requisite societal behavioural changes with the prognostic knowledge concerning the conditions and options for future scenarios of stable sustainability.

This morphological study identifies and measures recent nationwide trends in American street network design. Historically, orthogonal street grids provided the interconnectivity and density that researchers identify as important factors for reducing vehicular travel and emissions and increasing road safety and physical activity. During the 20th century, griddedness declined in planning practice alongside declines in urban form compactness, density, and connectivity as urbanization sprawled around automobile dependence. But less is known about comprehensive empirical trends across US neighborhoods, especially in recent years. This study uses public and open data to examine tract-level street networks across the entire US. It develops theoretical and measurement frameworks for a quality of street networks defined here as griddedness. It measures how griddedness, orientation order, straightness, 4-way intersections, and intersection density declined from 1940 through the 1990s while dead-ends and block lengths increased. However, since 2000, these trends have rebounded, shifting back toward historical design patterns. Yet, despite this rebound, when controlling for topography and built environment factors all decades post-1939 are associated with lower griddedness than pre-1940. Higher griddedness is associated with less car ownership - which itself has a well-established relationship with vehicle kilometers traveled and greenhouse gas emissions - while controlling for density, home and household size, income, jobs proximity, street network grain, and local topography. Interconnected grid-like street networks offer practitioners an important tool for curbing car dependence and emissions. Once established, street patterns determine urban spatial structure for centuries, so proactive planning is essential.

We consider the problem of maximizing portfolio value when an agent has a subjective view on asset value which differs from the traded market price. The agent's trades will have a price impact which affect the price at which the asset is traded. In addition to the agent's trades affecting the market price, the agent may change his view on the asset's value if its difference from the market price persists. We also consider a situation of several agents interacting and trading simultaneously when they have a subjective view on the asset value. Two cases of the subjective views of agents are considered, one in which they all share the same information, and one in which they all have an individual signal correlated with price innovations. To study the large agent problem we take a mean-field game approach which remains tractable. After classifying the mean-field equilibrium we compute the cross-sectional distribution of agents' inventories and the dependence of price distribution on the amount of shared information among the agents.

One of the major characteristics of financial time series is that they contain a large amount of non-stationary noise, which is challenging for deep neural networks. People normally use various features to address this problem. However, the performance of these features depends on the choice of hyper-parameters. In this paper, we propose to use neural networks to represent these indicators and train a large network constructed of smaller networks as feature layers to fine-tune the prior knowledge represented by the indicators. During back propagation, prior knowledge is transferred from human logic to machine logic via gradient descent. Prior knowledge is the deep belief of neural network and teaches the network to not be affected by non-stationary noise. Moreover, co-distillation is applied to distill the structure into a much smaller size to reduce redundant features and the risk of overfitting. In addition, the decisions of the smaller networks in terms of gradient descent are more robust and cautious than those of large networks. In numerical experiments, we find that our algorithm is faster and more accurate than traditional methods on real financial datasets. We also conduct experiments to verify and comprehend the method.

We continue a series of papers devoted to construction of semi-analytic solutions for barrier options. These options are written on underlying following some simple one-factor diffusion model, but all the parameters of the model as well as the barriers are time-dependent. We managed to show that these solutions are systematically more efficient for pricing and calibration than, eg., the corresponding finite-difference solvers. In this paper we extend this technique to pricing double barrier options and present two approaches to solving it: the General Integral transform method and the Heat Potential method. Our results confirm that for double barrier options these semi-analytic techniques are also more efficient than the traditional numerical methods used to solve this type of problems.

After the U.S market earned strong returns in 2003, day trading made a comeback and once again became a popular trading method among traders. Although there is no comprehensive empirical evidence available to answer the question do individual day traders make money, there is a number of studies that point out that only few are able to consistently earn profits sufficient to cover transaction costs and thus make money. The day trading concept of buying and selling stocks on margin alone suggests that it is more risky than the usual going long way of making profit. This paper offers a new approach to day trading, an approach that eliminates some of the risks of day trading through specialization. The concept is that the trader should specialize himself in just one (blue chip) stock and use existing day trading techniques (trend following, playing news, range trading, scalping, technical analysis, covering spreads) to make money.

The current knowledge system of macroeconomics is built on interactions among a small number of variables, since traditional macroeconomic models can mostly handle a handful of inputs. Recent work using big data suggests that a much larger number of variables are active in driving the dynamics of the aggregate economy. In this paper, we introduce a knowledge graph (KG) that consists of not only linkages between traditional economic variables but also new alternative big data variables. We extract these new variables and the linkages by applying advanced natural language processing (NLP) tools on the massive textual data of academic literature and research reports. As one example of the potential applications, we use it as the prior knowledge to select variables for economic forecasting models in macroeconomics. Compared to statistical variable selection methods, KG-based methods achieve significantly higher forecasting accuracy, especially for long run forecasts.

A theoretical method is empirically illustrated in finding the best time to forsake a loan such that the overall credit loss is minimised. This is predicated by forecasting the future cash flows of a loan portfolio up to the contractual term, as a remedy to the inherent right-censoring of real-world `incomplete' portfolios. Two techniques, a simple probabilistic model as well as an eight-state Markov chain, are used to forecast these cash flows independently. We train both techniques from different segments within residential mortgage data, provided by a large South African bank, as part of a comparative experimental framework. As a result, the recovery decision's implied timing is empirically illustrated as a multi-period optimisation problem across uncertain cash flows and competing costs. Using a delinquency measure as a central criterion, our procedure helps to find a loss-optimal threshold at which loan recovery should ideally occur for a given portfolio. Furthermore, both the portfolio's historical risk profile and forecasting thereof are shown to influence the timing of the recovery decision. This work can therefore facilitate the revision of relevant bank policies or strategies towards optimising the loan collections process, especially that of secured lending.

As a key indicator of unsafe driving, driving volatility characterizes the variations in microscopic driving decisions. This study characterizes volatility in longitudinal and lateral driving decisions and examines the links between driving volatility in time to collision and crash injury severity. By using a unique real-world naturalistic driving database from the 2nd Strategic Highway Research Program (SHRP), a test set of 671 crash events featuring around 0.2 million temporal samples of real world driving are analyzed. Based on different driving performance measures, 16 different volatility indices are created. To explore the relationships between crash-injury severity outcomes and driving volatility, the volatility indices are then linked with individual crash events including information on crash severity, drivers' pre crash maneuvers and behaviors, secondary tasks and durations, and other factors. As driving volatility prior to crash involvement can have different components, an indepth analysis is conducted using the aggregate as well as segmented (based on time to collision) real world driving data. To account for the issues of observed and unobserved heterogeneity, fixed and random parameter logit models with heterogeneity in parameter means and variances are estimated. The empirical results offer important insights regarding how driving volatility in time to collision relates to crash severity outcomes. Overall, statistically significant positive correlations are found between the aggregate (as well as segmented) volatility measures and crash severity outcomes. The findings suggest that greater driving volatility (both in longitudinal and lateral direction) in time to collision increases the likelihood of police reportable or most severe crash events... ...

This paper expands the notion of robust moment problems to incorporate distributional ambiguity using Wasserstein distance as the ambiguity measure. The classical Chebyshev-Cantelli (zeroth partial moment) inequalities, Scarf and Lo (first partial moment) bounds, and semideviation (second partial moment) in one dimension are investigated. The infinite dimensional primal problems are formulated and the simpler finite dimensional dual problems are derived. A principal motivating question is how does data-driven distributional ambiguity affect the moment bounds. Towards answering this question, some theory is developed and computational experiments are conducted for specific problem instances in inventory control and portfolio management. Finally some open questions and suggestions for future research are discussed.

AI systems have found a wide range of application areas in financial services. Their involvement in broader and increasingly critical decisions has escalated the need for compliance and effective model governance. Current governance practices have evolved from more traditional financial applications and modeling frameworks. They often struggle with the fundamental differences in AI characteristics such as uncertainty in the assumptions, and the lack of explicit programming. AI model governance frequently involves complex review flows and relies heavily on manual steps. As a result, it faces serious challenges in effectiveness, cost, complexity, and speed. Furthermore, the unprecedented rate of growth in the AI model complexity raises questions on the sustainability of the current practices. This paper focuses on the challenges of AI model governance in the financial services industry. As a part of the outlook, we present a system-level framework towards increased self-regulation for robustness and compliance. This approach aims to enable potential solution opportunities through increased automation and the integration of monitoring, management, and mitigation capabilities. The proposed framework also provides model governance and risk management improved capabilities to manage model risk during deployment.

Sex differences in early age mortality have been explained in prior literature by differences in biological make-up and gender discrimination in the allocation of household resources. Studies estimating the effects of these factors have generally assumed that offspring sex ratio is random, which is implausible in view of recent evidence that the sex of a child is partly determined by prenatal environmental factors. These factors may also affect child health and survival in utero or after birth, which implies that conventional approaches to explaining sex differences in mortality are likely to yield biased estimates. We propose a methodology for decomposing these differences into the effects of prenatal environment, child biology, and parental preferences. Using a large sample of twins, we compare mortality rates in male-female twin pairs in India, a region known for discriminating against daughters, and sub-Saharan Africa, a region where sons and daughters are thought to be valued by their parents about equally. We find that: (1) prenatal environment positively affects the mortality of male children; (2) biological make-up of the latter contributes to their excess mortality, but its effect has been previously overestimated; and (3) parental discrimination against female children in India negatively affects their survival; but failure to control for the effects of prenatal and biological factors leads conventional approaches to underestimating its effect by 237 percent during infancy, and 44 percent during childhood.

I characterize the consumer-optimal market segmentation in competitive markets with differentiated products. I show that this segmentation is public---in that each firm observes the same market segments---and takes a simple form: in each market segment, there is a dominant firm favored by all consumers in that segment. By segmenting the market, all but the dominant firm maximally compete to poach the consumer's business, setting price to equal marginal cost. Information, thus, is being used to amplify competition. This segmentation simultaneously generates an efficient allocation and delivers to each firm its minimax profit.

In the single IV model, current practice relies on the first-stage F exceeding some threshold (e.g., 10) as a criterion for trusting t-ratio inferences, even though this yields an anti-conservative test. We show that a true 5 percent test instead requires an F greater than 104.7. Maintaining 10 as a threshold requires replacing the critical value 1.96 with 3.43. We re-examine 57 AER papers and find that corrected inference causes half of the initially presumed statistically significant results to be insignificant. We introduce a more powerful test, the tF procedure, which provides F-dependent adjusted t-ratio critical values.