Inquire Europe’s 2019 autumn seminar in Krakow, Poland, brought together some 100 investment professionals and academics to discuss recent advances in “Investing with Machine Learning and New Techniques”.
Keynote address: machine learning
There is a lot of media chatter about artificial intelligence (AI) and how it is set to change our lives. Machine learning (ML) techniques, the most successful and promising branch of AI, excel at discovering non-linear relations in large amounts of data. How successfully ML techniques can be applied in finance was the topic of the Inquire Europe keynote address by Stefan Nagel, Professor of Finance at the University of Chicago and Executive Editor of The Journal of Finance.
Nagel argues that ML techniques are less well-suited than econometrics for hypothesis testing. While ML techniques and traditional econometrics share much in common, ML focuses more on out-of-sample prediction than on hypothesis testing. It centres around practical algorithms rather than theoretical statistical foundations and it makes minimal distributional assumptions.
Another point Nagel makes is that standard ML methods applied by the tech giants cannot be directly used in asset pricing. Among the special features of asset pricing models are low signal-to-noise ratios, changing relations and small databases in the time dimension (in the cross-sectional dimension the available data can be large).
Additionally, there is the danger of overfitting. Stefan Nagel argues that this can be managed by incorporating a priori knowledge. In asset pricing, there is the conventional wisdom, for instance, that individual factors/signals can only have modest price impacts (principle of shrinkage towards zero) and that the number of factors jointly predicting prices is likely to be small (principle of sparsity).
All-in-all, Nagel is cautious about expecting too much from advanced ML methods. Regarding overfitting, he argues that more traditional approaches such as principle component analysis (PCA) and regularised (constrained) regression techniques, like RIDGE and LASSO, are potentially well-suited to produce results in line with the principles of shrinkage and sparsity.
The other presentations discussed the usage of existing factors, see below, and the creation of new factors, discussed in Part 2.
Existing factors: Market timing
The presentations by Daniele Bianchi (Queen Mary University of London) and Dominik Wolff (DekaBank, Frankfurt) had a lot in common. Both dealt with short-term predictions culminating in active trading strategies. Both also considered various prediction models, ranging from the more conventional linear regression approaches such as PCA, RIDGE and LASSO, to today’s hot-topic non-linear machine learning approaches of decision trees (such as boosted trees and random forests) and neural networks.
Daniele Bianchi’s presentation focused on the prediction of returns on US Treasury bonds with different maturities. He examined two sets of prediction variables: a narrow set, containing only forward rates embedded in the yield curve, and a broader set with an extra 128 macroeconomic/financial variables. Non-linearity matters in the sense that neural networks deliver the best predictions and the extra variables helped to obtain better predictions than did forward rates alone. Typical macroeconomic variables such as inflation and output growth are important for predicting the long end of the yield curve, whereas financial variables, for example, the S&P 500 equity index and the credit spread, matter more for the short end.
Wolff selected 28 factors to predict the S&P 500. These included fundamentals (dividends/earnings), macroeconomics, sentiment and risks. Evaluation of the prediction errors showed good results for most of the analysed linear regressions. The tree-based non-linear ML techniques were less effective, however. According to Wolff, this was possibly because of the relatively short sample period and the low dimensionality of the problem (prediction of one index with 28 predictors).
Existing factors: Hedging
Emmanuel Jurczenko (GLION Institute, Geneva) discussed his paper on hedging macroeconomic risks. He pointed out that macroeconomic data is often noisy and frequently revised, thus leading to measurement errors and omitted-variable biases. Furthermore, directly trading macroeconomic indicators is typically not possible.
His proposed hedging solution takes these issues into account. It uses asset classes (equities, Treasuries, credits, inflation-linked bonds, commodities and FX baskets) as the building blocks of factor-mimicking portfolios (FMPs). The procedure for constructing these portfolios customises a relatively new technique. Jurczenko uses a sequence of linear regressions – subsequently LASSO, PCA, LASSO, and ordinary least squares (OLS). This ultimately results in the asset class loadings needed to construct the (rotated) FMPs. A typical US endowment portfolio served as a practical illustration in which growth, inflation and financial stress risks are hedged away.
The presentations summarised here examined the use of ‘existing’ factors. My main takeaways from them were:
- Machine learning excels at finding non-linear relations in big data sets. It uses practical algorithms that require only minimal distributional assumptions.
- Low signal-to-noise ratios, changing relations over time and lack of long data time series hinder the discovery of non-linear patterns in pricing data.
- Simplification in the estimation of linear structures is in practice often possible via PCA and/or constrained linear regressions. Jurczenko’s work on hedging macroeconomic risks in a diversified portfolio serves as an example.
- Bianchi argues that ML techniques do add value in predicting bond returns. In contrast, Wolff was unable to show this for the S&P 500.
For more articles on quantitative investing, click here >
For more articles by Erik Kroon, click here >
To discover our funds and select the ones that meet your requirements, click here >
Any views expressed here are those of the author as of the date of publication, are based on available information, and are subject to change without notice. Individual portfolio management teams may hold different views and may take different investment decisions for different clients.