This is the last issue of 2019 for the inaugural volume of JFDS. We have received wonderful feedback about the quality of the articles published in the first three issues from both the investment management and data science communities. This issue has nine articles, which I have summarized below.
For security analysts and asset managers who are accustomed to working with more established analytical frameworks and quantitative tools, financial data science must successfully demonstrate its applicability to them in the performance of their duties. As in any new research program that will go through various stages in order to become an industry standard, the early stages are especially important. It is in the early stages of financial data science where much of the field’s terminology, concepts, methodological foundation, and theory will be laid. Mathematics offers an example of a successful scientific endeavor of a research program. Drawing on the history of mathematics, Joseph Simonian in his article, “Proofs and Cross-Validations: Three Lessons for Financial Data Science,” provides three lessons for investment practitioners as to how financial data science has the potential to assist them in aligning their research priorities more closely with those of mainstream finance so as to produce investment research with lasting insight on the application of data science to finance. According to Simonian, financial data science must (1) be in epistemic dialogue with traditional finance, (2) aim for epistemic transparency, and (3) aim for epistemic connectivity. He concludes the article with some guidance on the related topic of effectively writing about and presenting financial data science research.
Empirical studies of numerous equity momentum strategies suggest that they offer a risk premium. These strategies require the explicit definition of both a trend estimator and a position-sizing rule. In “Enhancing Time-Series Momentum Strategies Using Deep Neural Networks,” Bryan Lim, Stefan Zohren, and Stephen Roberts introduce deep momentum networks, which are a hybrid class of deep learning models, that retain the volatility scaling framework of time-series momentum strategies while using deep neural networks to output position targeting trading signals. The authors evaluate two approaches to position generation. In the first approach, trend estimation is cast as a standard supervised learning problem by employing machine learning models to forecast the expected asset returns or probability of a positive return at the next time step and applying a simple maximum long–short trading rule based on the direction of the next return. In the second approach, which is based on trading rules directly generated from the model’s outputs, the authors calibrate by maximizing the Sharpe ratio or average strategy return. A universe of 88 continuous futures contracts is used to test the two approaches. Lim, Zohren, and Roberts report clear improvements in risk-adjusted performance by calibrating models with the Sharpe ratio. More specifically, they find that in the absence of transaction costs, the Sharpe-optimized long short-term memory model improved traditional methods by more than two times. When transaction costs of 2–3 basis points are considered, it continued outperforming traditional methods. In the case of more illiquid assets, the authors propose a turnover regularization term that trains the network to account for costs at run-time.
A neural network is an advanced optimization tool that has been studied in finance for more than 25 years and used in the 1990s for pricing derivatives. Following a trial-and-error process, the output of a neural network is complex functional relationships between a set of observable inputs and outputs. Despite its interest by financial modelers, there have been few real-world applications to the various activities of asset management. In “Neural Networks in Finance: Design and Performance,” Irene Aldridge and Marco Avellaneda first describe a step-by-step technique for building a potentially profitable financial neural network, explaining the various components of neural networks and comparing popular neural network activation functions and their applicability to financial time series. They find that the tanh activation function closely mimics financial returns and produces the best results. They also explain the relationship between neural network modeling and traditional forecasting and econometric modeling. Using daily and monthly financial data, the authors then provide an application of a neural network to investing. More specifically, Aldridge and Avellaneda find that neural network-based daily trading strategies on major US stocks significantly and consistently outperform a buy-and-hold position in the same stocks. The article provides a guide for practitioners in designing systems that can be used to capitalize on market inefficiencies.
Systematic trading strategies involve the constructing of portfolios that are expected to generate optimized performance based on a set of rules. To attain any of the desired return profiles quantitative asset managers must determine a large array of trading parameters, which include the candidate of assets that may be included in the portfolio, the appropriate trading horizon, the type of strategy being considered, and many other possible parametrizations. Typically, in analyzing a proposed systematic trading strategy, an asset manager attempts to identify the appropriate parameters by either using all the historical data available or holding out some of the most recent data for use as an out-of-sample validation set. There are drawbacks in backtesting that can render the results misleading regardless of whether the full or holdout sample for validation are used. Three proposals for overcoming the deficiencies in backtesting suggested in the literature are data snooping, overestimated performance, and cross-validation evaluation. In “Avoiding Backtesting Overfitting by Covariance-Penalties: An Empirical Investigation of the Ordinary and Total Least Squares Cases,” Adriano Koshiyama and Nick Firoozye propose a new approach to dealing with financial overfitting: covariance-penalty correction, which corrects the Sharpe ratios based on the number of parameters in a model. Using ordinary and total least squares, the authors undertake an empirical investigation and compare the performance of the covariance-penalty correction method with some other approaches in the realm of covariance-penalties across more than 1,300 assets. The results reported by Koshiyama and Firoozye suggest that the covariance-penalty correction approach is a suitable procedure to avoid backtesting overfitting, and total least squares provides superior performance when compared to ordinary least squares.
There are several methods proposed in the investment literature for differentiating between emerging and developed markets. In practice, index providers differ in their classification methodologies for both their equity and bond indexes. In the case of bond indexes, there is even greater variation used by index providers. Classification of developed and emerging markets is not a mere academic exercise. It is critically important for both active and index investors. The article “Reconstructing Emerging and Developed Markets Using Hierarchical Clustering” describes how unsupervised machine learning techniques can be employed to objectively group countries from an investment perspective. Going beyond classifications based on economic fundamentals, the authors, Gerald Garvey and Ananth Madhavan, group countries based on returns in equity and bond markets. There are two ways in which their analysis is distinctive from the literature. First, they focus exclusively on the problem of grouping countries without taking a stance on which countries or groups are better (either as tactical investments as done in the practitioner literature or as a guide to how economies should be managed as characterized in the academic literature). Second, by examining both stock and bond markets, Garvey and Madhavan gain insight into how to construct investible vehicles that better capture emerging market exposure and alpha. They find an important geographical footprint that differs significantly from the groupings that are used by most practitioners.
It is well recognized that it is critical to consider tail events when backtesting risk premium strategies using historical data. However, because such tests include few observations that are tail events, it is difficult for a backtest to evaluate both the true risks involved in a strategy and the probability that a strategy will outperform a benchmark in the long term. Although the literature suggests employing bootstrapping procedures to overcome these backtesting problems, these procedures are not ideal for two reasons. The first drawback is that they re-employ the same extreme residuals repeatedly. The second drawback is that they do not simultaneously preserve both the time-series and cross-sectional dependencies of returns. Yet it is from the performance from both time-series and cross-sectional dependencies that many cross-asset class strategies, such as time-series momentum, draw their performance. In “Time-Series Momentum: A Monte Carlo Approach,” Clemens Struck and Enoch Cheng propose a Monte Carlo procedure that addresses these two drawbacks. By using the empirical residual distributions to generate a large variety of new residuals instead of resampling, the first drawback is addressed; by combining time-series models and copulas, the second drawback is addressed. The authors generate 10,000 paths of different time-series momentum strategies based on the S&P 500 and a cross-asset-class futures portfolio. The simulation results indicate that the probability distribution for the strategies that do outperform buy-and-hold in-sample using historical backtests may (1) exhibit tail risks and (2) underperform or outperform when out-of-sample. Struck and Cheng report that the findings are unchanged when employing different time-series momentum strategies, time periods, asset classes, and risk measures.
A major concern with algorithmic trading is that it may amplify the price volatility of a traded asset that, in turn, can lead to a cascading of that asset’s price. The Flash Crash of May 6, 2010 is an example. Jung Heon Song, Marcos López de Prado, Horst D. Simon, and Kesheng Wu in their article “Extracting Signals from High-Frequency Trading with Digital Signal Processing Tools” address this issue by exploring a number of signal processing tools to understand the trading of natural gas (NG) futures and the impact of temperature forecasts on the volatility of trading. (For many years, NG futures have been one of the most heavily traded energy contracts.) The authors demonstrate that signal processing tools can effectively extract useful patterns from the trading records of NG futures trading data. They employ three sets of techniques: Fourier analysis, Lomb–Scargle periodogram, and cointegration. Fourier analysis of NG futures trading records indicates that (1) the high-frequency components have been increasing over the past few years, (2) there is a strong once-per-minute signal, (3) the presence of time-based algorithmic traders, and (4) there are many more trades occurring during the first second of a minute, which also indicates automated trades triggered by the clock. Using the technique of cointegration, the authors investigate the relationship between temperature forecasts and NG futures prices. Their findings indicate that the two are cointegrated when the trading data are separated into seasons. Furthermore, analyzing the role the Global Ensemble Forecast System (GEFS)—created by the National Centers for Environmental Prediction to address the uncertain nature of numerical weather prediction—they estimate that weather forecasting errors are the reason for the significant percentage of the daily volatility in NG prices, even for the most accurate GEFS model they study. Thus, even a relatively small forecast temperature error could propagate into a larger change in the price of NG.
Although there are some differences, in the cryptocurrency space an initial coin offering (ICO) is fundamentally similar to an initial public offering (IPO). In an ICO offering document, there are different levels of information disclosed: team, whitepaper, code and/or prototype, roadmap, and token sales terms. It would be beneficial for potential investors to use ICO disclosure to segregate ICOs into different topics. By doing so, potential investors can decide which ICO they should focus on for further analysis. Potential investors could be interested in topics (e.g., finance, media, information, professional services, health and social, natural resources). Fu Chuanjie, Andrew Koh, and Paul Griffin, in their article “Automated Theme Search in ICO Whitepapers” apply the latent Dirichlet allocation (LDA) model for building an automated tool that differentiates ICO whitepapers according to different topics, themes, domains, or industries. (Whitepapers that are included in the ICO disclosure document highlight the technology proposal.) This is done by fitting the LDA model to the text extracted from the ICO whitepapers. They evaluate the automated categorization of whitepapers using two methods: statistical methods and human judgment methods. The authors find that there is enough evidence to conclude that the LDA model appropriately categorizes the ICO whitepapers. Moreover, Chuanjie, Koh, and Griffin find a statistically significant difference between topics in the success of an ICO being funded. This suggests that the topics are usefully differentiated and that their proposed topic model could be used to help predict whether an ICO will be successful.
Section 13f of the Securities Exchange Act passed in 1975 and mandated that large institutional investors (i.e., those managing more than $100 million) disclose portfolio holdings each quarter. The information provided by 13f filings offers all asset managers, as well as individual investors, the opportunity to construct portfolios like those of the reporting institutional investors, a process referred to as alpha cloning. However, alpha cloning is far from a simple process because the holdings may not be representative of the long-term holdings of reporting managers and it is difficult to distinguish positions that represent opportunistic holdings from those that are for diversification and stability purposes. In “Alpha Cloning: Using Quantitative Techniques and SEC 13f Data for Equity Portfolio Optimization and Generation,” Daniel M. DiPietro investigates several quantitative techniques for portfolio optimization and generation using these filings. He finds that by using the Sharpe ratio, optimized portfolios based on 13f data outperformed a traditional market capitalization (approximately 2 versus 1.29); however, based on returns, optimized portfolios did not outperform a traditional market-capitalization portfolio. The results were mixed for portfolio generation models. Although some of the portfolios failed to break even, others outperformed the S&P 500 by a large margin. The firm-specific overweighted investment extraction with market cap balancing was the most successful generation model, consistently generating more than 85% in returns over the five-year backtesting period, even surpassing 95% for various numbers of holdings. S&P 500, in contrast, yielded approximately 72% over the same period. Overall, the results reported in this article suggest that SEC 13f data have the potential to build long-term equity quantitative models that achieve noteworthy performance.
Francesco A. Fabozzi
Managing Editor
- © 2019 Pageant Media Ltd