Systematic macro investors’ use of the regime-switching models that have been developed in academia over the last several decades is infrequent at best and, when used, generally tangential to their core investment process. The roots of this less-than-enthusiastic uptake can be found in two familiar sources: models that possess an overly complex formal structure and poor predictive ability. As a remedy to the current state of affairs, the authors present a new foundation for regime-based investing, one based on spectral clustering, a graph theoretic approach to classifying data. Drawing inspiration from the work of Hyman Minsky and John Geanakoplos, the authors present a macro framework that uses measures of growth, inflation, and leverage to define regimes and drive portfolio decisions. To the latter end, the authors show how the framework can be used to build portfolios using information about regimes as defined, to outperform a no-information equal-weight portfolio both out-of-sample and in bootstrapped and cross-validated simulations. The authors thus show that spectral clustering can provide both an elegant mathematical description of the leverage cycle and a robust foundation for quant-macro investing.

Individual analysts typically publish recommendations several times per year on the handful of stocks they follow within their specialized fields. How should investors interpret this information? How can they factor in the past performance of individual analysts when assessing whether to invest long or short in a stock? This is a complicated problem to model quantitatively: There are thousands of individual analysts, each of whom follows only a small subset of the thousands of stocks available for investment. Overcoming this inherent sparsity naturally raises the question of how to learn an analyst’s forecasting ability by integrating track-record information from all the stocks the analyst follows; in other words, inferring an analyst’s ability on Stock X from track records on both Stock X and stocks other than X. The authors address this topic using a state-of-the-art computationally rapid Bayesian machine learning technique called independent Bayesian classifier combination (IBCC), which has been deployed in the physical and biological sciences. The authors argue that there are many similarities between the analyst forecasting problem and a very successful application of IBCC in astronomy, a study in which it dominates heuristic alternatives including simple or weighted averages and majority voting. The IBCC technique is ideally suited to this particularly sparse problem, enabling computationally efficient inference, dynamic tracking of analyst performance through time, and real-time online forecasting. The results suggest the IBCC technique holds promise in extracting information that can be deployed in active discretionary and quantitative investment management. <b>TOPICS:</b>Security analysis and valuation, big data/machine learning, equity portfolio management

In the classic mean-variance portfolio theory as proposed by Harry Markowitz, the weights of the optimized portfolios are directly proportional to the inverse of the asset correlation matrix. However, most contemporary portfolio optimization research focuses on optimizing the correlation matrix itself, and not its inverse. In this article, the author demonstrates that this is a mistake. Specifically, from the Big Data perspective, she proves that the inverse of the correlation matrix is much more unstable and sensitive to random perturbations than is the correlation matrix itself. As such, optimization of the inverse of the correlation matrix adds more value to optimal portfolio selection than does optimization of the correlation matrix. The author further shows the empirical results of portfolio reallocation under different common portfolio composition scenarios. The technique outperforms traditional portfolio allocation techniques out of sample, delivering nearly 400% improvement over the equally weighted allocation over a 20-year investment period on the S&P 500 portfolio with monthly reallocation. In general, the author demonstrates that the correlation inverse optimization proposed in this article significantly outperforms the other core portfolio allocation strategies, such as equally weighted portfolios, vanilla mean-variance optimization, and techniques based on the spectral decomposition of the correlation matrix. The results presented in this article are novel in the data science space, extend far beyond financial data, and are applicable to any data correlation matrixes and their inverses, whether in advertising, healthcare, or genomics.