Advancing Asset Selection and Portfolio Design with Machine Learning Methods

In today’s world of asset selection and portfolio management, one of the most difficult parts of separating oneself from the rest of the pack is finding novel methods for picking stocks and building successful strategies in order to gain a competitive advantage.

Conventional methods can be, and are currently, used by managers to build portfolios and attempt to optimize asset weighting to achieve their desired outcome. Yet with well-understood and widespread strategies come the limitations of such methods. The asset management space is saturated with these conventional strategies, many of which have existed for decades.

It is difficult for managers to find a way to put themselves ahead of competitors. Arbitrage opportunities get smaller and the margins thinner as more people take advantage of well-known asset selection models.

Modern Machine Learning Methods and Value

New machine learning (ML) models allow managers navigate the waters with more confidence and find opportunities in the market that may go undetected by conventional methodology.

LASSO Regression

LASSO regression was introduced by Robert Tibishrani in 1996 (“Regression Shrinkage and Selection via the Lasso”. Journal of the Royal Statistical Society. Series B (Methodological), vol. 58, no. 1, 1996, pp.267-288.) as a way to perform both variable selection and regularization so as to enhance the prediction accuracy and interpretability of the statistical model it produces.

Many managers still employ more basic statistical analysis to select the best assets that have a chance of future profitability, but there are considerably larger margins for error and variance in the outcomes of portfolios created this way.

LASSO can allow a manager to process larger data sets and more accurately select a subset of assets amongst thousands which have highest chance of future profitability while limiting exposure to the risks faced by their benchmark index.

Ridge Regression

Using ridge regression, as introduced by Hoerl and Kennard in 1970 (“Ridge Regression: Biased Estimations for Nonorthogonal Problems”. Technometrics, vol. 12, no. 1, 1970, pp. 55-67.), it is possible to build upon the performance of the mean-variance problem outlined in traditional Markowitz portfolio theory (“Portfolio Selection”. The Journal of Finance, vol. 7, no. 1, 1952, pp. 77-91.)

This method allows managers to avoid over-fitting their models – overfit models deliver superb back test results but are often a poor predictor of future performance.

The accuracy of a model relying on simple linear regression methods can fail when processing a large set of assets by improperly identify correlation between asset movements, either up or down. Ridge regression, on the other hand, helps sort through the noise and find the assets or factors which will allow a portfolio model to perform optimally.

Benefits of Machine Learning Models

In today’s market environment, asset selection and portfolio optimization can be limited by physical time and data constraints. Every day there is more data available for managers to use, but the physical time constraints of analysts and computational power of conventional data science are unable to fully leverage the available data.

Millions of data points that were previously inaccessible to financial analysts are now available to those who employ ML strategies. Managers no longer need to prioritize the testing of new strategies based on current human or data science limitations – ML techniques can automate many of the processes involved in portfolio management and can allow managers to better allocate their existing resources.

Portfolios such as the Kai Enhanced Sector Portfolio employ ML to give managers access to market sector returns while minimizing sector risk by picking assets with high upside potential and avoiding exposure to assets that look unlikely to produce positive returns.

By leveraging the suitable ML techniques, managers can build smarter portfolio models that learn and evolve as their data feed grows, and their models can home in on areas in the market that provide alpha where conventional strategies might have overlooked.

Machine Learning in Industry Practice

Portfolio selection that is informed by market sentiment and/or macro-economic activity can employ natural language processes (NLP) to analyze and synthesize market news and information in an automated way that previously relied on human analysts. NLP provides a constant, streamlined source of information that allows analysts to better locate pertinent market information to inform decisions when selecting assets or developing portfolio management strategies.

BlackRock has launched a new line of actively managed exchange-traded funds (ETFs) which employ NLP to examine regulatory filings, corporate earnings reports, and other market news to determine which assets to hold and how the composition of the ETFs will change. Traditional sector ETFs rely on historical industry classification, but BlackRock’s NLP technology enables them to reclassify company industries as they evolve in order to build more accurate, forward-looking sector ETFs. Using machine learning techniques to manage and monitor their new line of ETFs has allowed BlackRock to lower its fees in order to entice more customers.

Putnam Investments, a provider of mutual funds, institutional investment strategies and retirement services, has embraced ML and AI as part of their future. The firm has used machine learning to gain insights from enormous data sets that were previously inaccessible, and are building ML algorithms to make smarter investment product recommendations.

Portfolio selection methods and conventional wisdom will not die out, but machine learning can help portfolio managers analyze and synthesize more data than ever before and use this information to better inform their decision-making process. No matter what aspect of asset selection or management you are involved in, machine learning techniques will provide a competitive edge in a crowded field.

Hoerl, Arthur E. and Robert W. Kennard. “Ridge Regression: Biased Estimations for Nonorthogonal Problems”. Technometrics, vol. 12, no. 1, 1970, pp. 55-67.

Markowitz, Harry. “Portfolio Selection”. The Journal of Finance, vol. 7, no. 1, 1952, pp. 77-91.

Tibshirani, Robert. “Regression Shrinkage and Selection via the Lasso”. Journal of the Royal Statistical Society. Series B (Methodological), vol. 58, no. 1, 1996, pp.267-288.