The authors use neural networks to examine the power of Treasury term spreads and other macro-financial variables to forecast US recessions and compare them with probit regression. They propose a novel three-step econometric method for cross-validating and conducting statistical inference on machine learning classifiers and explaining forecasts. They find that probit regression does not underperform a neural network classifier in the present application, which stands in contrast to a growing body of literature demonstrating that machine learning methods outperform alternative classification algorithms. That said, neural network classifiers do identify important features of the joint distribution of recession over term spreads and other macro-financial variables that probit regression cannot. The authors discuss some possible reasons for their results and use their procedure to study US recessions over the post-Volcker period, analyzing feature importance across business cycles. <b>TOPICS:</b>Fixed income and structured finance, big data/machine learning, financial crises and financial market history <b>Key Findings</b> ▪ It is difficult to make the case that neural network classifiers greatly outperform traditional econometric methods such as probit regression when forecasting US recessions if performance is measured on the basis of forecast accuracy alone. ▪ That said, neural network classifiers identify important features of the joint distribution of recession over term spreads and other macro-financial time series that probit regression and other traditional methods cannot. ▪ The authors propose a three-step econometric process for conducting statistical inference on machine learning classifiers, with the goal of better explaining recession forecasts and enabling model outputs to be linked back to the instruments of monetary policy.

The authors examine an optimized Markowitz efficient portfolio by applying a quantitative trading strategy to the S&amp;P 500 sector exchanged-traded funds (ETFs). First, they implement a pattern-matching trading system, which extracts the underlying trends based on dynamic time warping. They then estimate a decision-making dictionary from the windows of ETF prices to identify the entry points for trading. Finally, they construct a Markowitz efficient portfolio on the ETFs’ net asset values on the validation set. The results demonstrate that the strategy can be modified to improve performance. <b>TOPICS:</b>Exchange-traded funds and applications, portfolio construction, statistical methods <b>Key Findings</b> ▪ The authors explore the applicability of dynamic time warping to financial time series in the context of quantitative trading strategies. ▪ They construct a portfolio that minimizes the expected volatility by estimating the optimal weights for each component to explore profitable quantitative strategies. ▪ They demonstrate the flexibility of the pattern-matching trading strategy by modifying the strategies to adapt to both the pre–COVID-19 period and the post–COVID-19 period.

The dominant approaches for financial portfolio construction rely on estimating sample covariance and correlations matrixes, which serve as an input into a number of classical portfolio construction techniques. These classical approaches are not forward looking, are constrained by the ability to estimate covariance and correlation matrixes, and are inflexible to incorporating additional information. The authors propose a new approach of using learned representations from deep learning networks to augment such classical techniques. This approach can incorporate learned estimates of future performance and can be customized to create tailored representations best suited to meeting varying financial objectives. This article showcases one example of such an embedding, compares and contrasts it with classical approaches to portfolio construction, and discusses additional possibilities for applying representation learning in quantitative finance. TOPICS:Big data/machine learning, performance measurement, portfolio construction, simulations, statistical methods Key Findings • The authors propose a general methodology for creating tailored embeddings of financial time-series data. • The authors demonstrate one approach for creating such embeddings, analyze the induced similarity measure, and compare and contrast it with the traditional approach of using covariance and correlation as a similarity measure for securities. • The authors discuss other applications and advantages of this approach, including accounting for probabilistic estimates of future behavior and flexibility in incorporating varying business objectives.

As the aging baby boomer generation saves for retirement and demands suitable retirement savings vehicles, the role of asset and custody management banks is growing and the assets under their management are rising. The authors used a data envelopment analysis (DEA) model to evaluate the economies of scale in the asset management and custody bank subindustry of the capital market industry for the period of 2010 to 2017. They measured company size in two ways: total assets under management of the firm and total revenue generated by the firm. The authors evaluated cost efficiencies with cost measured in terms of total expenses, and they evaluated two important components of total expenses: total distribution expenses and general administrative and selling expenses. Using a returns to scale (RTS) DEA model, they found that most asset management firms initially exhibit increasing RTS until they become too large. The trend persisted when general and administrative expenses were evaluated and when total distribution expenses were evaluated. TOPICS:Portfolio management/multi-asset allocation, portfolio construction, wealth management, retirement Key Findings • Initially most asset management firms exhibit increasing returns to scale (IRS) until they become too large. • As companies expand in terms of total assets and total revenues, they do reach a point at which size is too large and cost inefficiencies manifest in the form of decreasing returns to scale. • Whenever asset management and custody banks reach their most productive scale size and begin to experience decreasing returns to scale, they make structural changes by adopting new technologies and/or changes in the processes so that their total operating expenses decline and they can grow again without a proportionate increase in operating costs.

Price movement predictions of financial instruments using traditional time-series models with a predefined mathematical structure are common, but this method restricts their ability to learn latent patterns in the data. In recent times, artificial neural networks (ANN) have been able to learn complex hidden patterns from financial datasets using a highly nonlinear architecture. However, most experiments with neural networks require a lot of time to search a suitable network and subsequently train the network. The authors have developed a deep multilayer perceptron (MLP) classifier with a zero-centered piecewise linear unit activation that yielded better classification performance according to their accuracy metric and required consistently less training time compared to a similar MLP network with leaky rectified linear unit and tanh activation functions. The authors illustrate their technique with a large high-frequency dataset on selected bank shares from the Indian stock market. The authors also discuss the theoretical properties and advantages of their proposal. <b>TOPICS:</b>Statistical methods, simulations, big data/machine learning <b>Key Findings</b> • Time-series models have been used extensively to understand stock price movements. They are constrained by a predefined structure. Artificial neural network (ANNs), due to their universal approximation properties, offer more flexibility; however, they suffer from long training times. • In this article, the authors introduce a zero-centered piecewise linear unit (PLU), a hybrid activation function derived from tanh and leaky rectified linear unit (ReLU) functions that exhibits the zero-centered property from tanh and the nonzero gradient property from leaky ReLU. • The authors developed a multilayer perceptron (MLP) classifier with PLU activation, which yielded better classification performance according to the accuracy metric and required consistently less training time compared to a similar MLP network with leaky ReLU and tanh activation.

The main goal of this article is to show the relation between global equity and bond funds from a network perspective. The authors demonstrate the advantages of graph theory to explain the collective fund dynamics. The results show that equity and bond funds have a significant exposure to the Fama–French–Carhart factors. The authors argue that the network is dynamically driven by equity funds with their centrality scores and risk factor exposure and can transmit and amplify system-wide stress or inefficiencies in the factor bets. Using graph theory, the authors demonstrate that the return-based relationships between bond and equity funds are asymmetrical and the network is sufficiently clustered. Specifically, equity funds connect the different clusters. The HML factor is significant both on a single-fund level and as a web determinant. Therefore, investors should pay close attention to it when managing funds and deriving asset allocations. Finally, the authors provide a machine learning approach to how fund managers, plan sponsors, and analysts can derive equity–bond allocations, based on centrality scores, factor exposure, and hierarchical clustering of asymmetrically connected assets. <b>TOPICS:</b>Equity portfolio management, statistical methods, simulations, big data/machine learning <b>Key Findings</b> • The authors use topology and a concept from physics to show strong–weak return- and factor-based relationships between global equity and bond funds using a directed network approach. • The Fama–French–Carhart factors determine an asymmetrical bond–equity fund relation and different fund clusters. The network exhibits a structure that differs according to the market cycle. Equity funds play the most central role in the market and connect the different clusters. • Investors can use the bond–equity interconnectedness and network topology in multiple ways. The authors show that simple passive allocations, fund-of-funds solutions, and cluster-wise control for risk factors are some of the possible applications.

Given only the historic net asset value of a large-cap mutual fund, which members of some universe of stocks are held by the fund? Discovering an exact solution is combinatorially intractable because there are, for example, C(500, 30) or 1.4 × 1048 possible portfolios of 30 stocks drawn from the S&P 500. The authors extend an existing linear clones approach and introduce a new sequential oscillating selection method to produce a computationally efficient inference. Such techniques could inform efforts to detect fund window dressing of disclosure statements or to adjust market positions in advance of major fund disclosure dates. The authors test the approach by tasking the algorithm with inferring the constituents of exchange-traded funds for which the components can be later examined. Depending on the details of the specific problem, the algorithm runs on consumer hardware in 8 to 15 seconds and identifies target portfolio constituents with an accuracy of 88.2% to 98.6%. TOPICS:Big data/machine learning, statistical methods, portfolio management/multi-asset allocation

Portfolio selection based on pattern matching has shown great potential. The authors show that this approach can be derived from a symmetric market perspective, in which the relationship between market status and optimal portfolio is quantitatively defined in terms of a Pearson correlation. This new perspective motivated a revised pattern-matching algorithm (symmetric CORN-K), which selects the portfolio that simultaneously maximizes the returns of similar periods and minimizes the returns of dissimilar periods. The algorithm was further extended to a general symmetry-based pattern-matching algorithm (functional CORN-K) that uses the symmetry property in a principled way. The authors’ experiments demonstrated that the new algorithms can deliver better returns, larger Sharpe ratios, and lower maximum drawdown, and that the improvements are statistically significant.

To price and risk-manage over-the-counter derivatives, financial institutions have to estimate counterparty default risks based on liquidly quoted credit default swap (CDS) rates. For the vast majority of counterparties, liquid CDS quotes are not available and proxy CDS rates need to be constructed. Existing methods ignore counterparty-specific default risks and can lead to arbitrage. The authors propose to construct CDS proxy rates by machine learning techniques to associate liquidly quoted CDS proxy names to illiquid ones on the basis of observable financial feature variables. A benchmarking exercise shows that the proposed method leads to significantly smaller estimation errors than do existing methods. The authors tested 126 classifiers coming from the eight most popular algorithms, rank-ordered performance, and investigated performance variations among and within the classifiers. In Part I, the authors present the methodology, review the different machine learning techniques, and report on cross-classifier performance. In Part II, they focus on parametrization and intra-classifier performance, investigate correlation effects, and perform a benchmarking exercise. This is a first systematic study of CDS proxy construction by machine learning techniques and a first classifier comparison study entirely based on financial market data. The techniques should be of interest for financial institutions seeking proxies for CDS rates or other financial variables.

This is the second of two articles by the authors on the construction of credit default swap (CDS) proxy rates. In the first article, the authors proposed a machine learning (ML)-based proxy-rate construction technique that uses classification to construct so-called proxy names whose liquidly quoted CDS rates can be used as CDS proxy rates. The authors then compared the performances of ML classifiers from the eight most popular classifier families as a function of carefully selected sets of feature variables. In this second article, the authors take a closer look at the performances of the individual classifiers as a function of their different parametrizations, which they refer to as an intra-classifier performance study. The authors also examine the effects of feature variable correlations on classifier performance and perform a benchmarking exercise by comparing the ML-based CDS-proxy technique with two currently used proxy-rate construction methods: curve mapping and cross-sectional regression.