Value, value and more value

Revisiting our quantitative approach to style bias prediction, we continue to find strong support for a value bias in Asian equities over the next 12 months. 98% of historic months with current macro regime characteristics saw Asian value beat Asian growth over the fol-lowing year. Key risk? US 10 year bond yields.

Value, value and more value

The fourth edition of this occasional series, released in November 2016 and titled Next steps in style and size, provided a selection of quantitative models to support our view that the macro backdrop was increasingly tilting in favour of the value style outperforming the growth style in Asia (and in the US).

The time frame built into that analysis was 12 months+. With the benefit of hindsight this call was early, as we saw the outperformance in the growth style strengthen in calendar 2017 (principally driven by the largest Asian tech stocks). From a macro perspective, we had expected US 10 year yields to move higher sooner, and these yields are a key driver of our style bias analysis.

Nonetheless, readers of both this series and our monthly reports for our Asian systematic equities fund will be aware that we have stuck with the view that the macro backdrop was continuing to head in a direction favouring value over growth. The longer the opposite persisted (as was the case in 2017), the more acute the expected backlash when the value bias finally materialized. Finally in 2018 we began to see the landscape shift in favour of value, especially in the second half of this year. This short note revisits our 2016 analysis, re-checks the model and assesses current signals.

Figure 1. 12 month forward Asian equities style bias (December 2000+)
Splitting criteria are applied to each monthly observation to divide the sample into periods in which growth outperformed value or value outperformed growth, e.g. US 10 year Government yield < 2.6%. If the splitting criteria is true, then move to the left on the tree, if false then move right. Numbers below each node reflect the number of monthly observations meeting the previous splitting criteria, split into periods in which growth outperformed value (left) or value outperformed growth (right). E.g. For terminal node 1 there were a total of 62 observations for which 10 year yields were less than 2.6% and the ISM was greater than 50, with 58 of these coinciding with growth stocks outperforming value stocks for the subsequent rolling 12 month period. 4 observations subsequently saw value outperform growth. The ratio of 58:4 in favour of growth therefore identifies this as a growth-biased node.

Our decision tree model continues to favour value over growth. Recent selling has had no impact on that call. This approach strongly points to a value bias in Asian stock performance over the next year.

Decision tree modeling

To paraphrase our 2016 paper, decision tree models employ quantitative techniques to identify the most critical factors to explain/predict economic regimes. Essentially, once fed a multitude of factors, the model selects the one critical question to ask regarding regime identification. Having identified that first factor (and thus creating two “branches”) the model then identifies the next best question to further differentiate between regimes, resulting in additional branches . This process continues until the model is unable to improve upon the solution without introducing additional error. The result is a decision tree – effectively a flow diagram – for identification and prediction of regimes.¹

The strength of this approach is its ability to identify the most useful decision factors in complex multi-dimensional problems. The key weakness is, if misused, this approach can be no more than useless data mining. It is therefore important that we are able to identify a sensible economic rationale for each branch/node of the tree. In addition, by evaluating different, but related decisions, across a range of time periods, we can look for common themes which would signal greater model robustness.

For each of these tests the model was provided with a wide range of common economic factors (including interest rate data, yield curve data, credit spreads, US business confidence, US market valuation and recession indicators). These factors were selected on the basis of a considerable body of academic research highlighting their importance for economic regime identification. All bond-related data are in yield terms. A description of each variable is provided in the appendix. Note it is remarkable that, for the Asia style tree, only three macroeconomic factors were required to identify major regimes (US 10 year yields, the Fed Funds rate and US manufacturing confidence).

The dependent variable in the test was the market style bias over the period 12 months forward from the explanatory factors (as defined by MSCI style indices), i.e. whether or not value had outperformed growth over the period 12 months after each factor data point.

Results

The resulting decision tree for Asian style bias is presented in Figure 1. This tree is virtually identical to that presented in November 2016, despite being updated with data since then, through to October 2018.

To interpret this diagram, begin at the top with the deciding factor being the US 10 year government bond yield. Historically, when this has been very low (<2.6%) and the ISM (NAPM PMI index) has been above 50, growth has outperformed value in the Asia over the subsequent 12 months (node 1). In this case there were 62 observations with these macro characteristics, and of those, 58 subsequent 12 month periods saw growth outperform value, while only 4 saw value outperform growth. Hence, approximately 94% of observations were growth biased, so we label this a growth node.

All other nodes can be read in a similar manner, with a “true” response to each decision going to the left and a “false” response to the right. We derive 7 nodes (3 growth, 4 value) representing style bias in each of these economic regimes.

Figure 2 provides a conceptual economic growth cycle and each node is mapped to a point in that cycle.

We can see that the growth style tends to dominate around mid/late cycle and at a cycle trough. Value tends to dominate very late cycle, at the cycle peak, during certain parts of cycle contractions and post cycle-recovery.

Given the increase in the US 10 year bond yield, we have this year moved from node 1 to node 4. Put simply, we have moved from late cycle to peaking cycle. Historically, when US 10 year yields were above 2.6%, the Fed could be raising rates but not especially aggressively, and the ISM was not shooting up, Asian value outperformed Asian growth. We have seen 68 months with this combination of characteristics since 2000. Of those, 67 saw value outperform growth over the next 12 months; a 98% success rate in regime identification.

What could trip this up? Bond yields. A large fall in bond yields (back towards yields in the mid 2’s or lower) would signal a return to a growth bias in stock selection. Our view is this is not particularly likely in the near term given, amongst other reasons, requirements for large increases in issuance on the back of US federal funding requirements (combined with existing private issuance demands). Indeed, there could be some unusual yield behavior over the next year or two (driven by private investment crowding out effects) with the potential for rising long yields even as the US economic peaks and begins to slow.

Appendix – Data definitions

Fed target rate
US Federal Reserve target rate (and upper bound). Changes are in basis point terms over the period referred to in the node
description.

US 10 year
US Government 10 year bond yield. Changes are in basis point terms over the period referred to in the node description.

ISM
Institute for Supply Management Purchasing Managers’ Index. Changes are in simple difference terms over the period referred to in the node description.

¹ Technically, we employ recursive partitioning using R Core Team (2015) – R. A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/

Written by: Dr. Hamish Macalister