This is the third installment of a series of posts highlighting the risk-measurement benefits of a model that links separate regional models, versus a single global model. In the first installment, “Capturing Regional Nuances with a Global Linked Model,” we took the point of view of a risk manager of a multimanager portfolio and to show how a linked model can highlight differences in regional exposures that might get lost in a single global model. In the second post, “A ‘Link’ to Better Global Risk Forecasts,” we illustrated the benefits to risk forecast accuracy of a linked model over a global model when optimizing a global Momentum portfolio. In this post we address another portfolio management strategy: allowing a portion of a portfolio to be invested outside the geographic region covered by the benchmark.
It is not unusual for a mandate with a developed market benchmark to allow a portion of the portfolio to invest in emerging markets. To ensure that all risks are covered in the risk forecasts, a manager would have to use a model that incorporates both developed and emerging markets. At Qontigo, that risk model might be the Worldwide (WW4) model, but it could also be the new Worldwide Equity Linked Factor Risk Model (LM4) that knits together our US model (US4), a Developed Markets ex-US model (DMxUS4) and an Emerging Markets model (EM4). Because factor returns and volatilities may vary, sometimes substantially, across regions, the model that recognizes those nuances should provide a better overall assessment of the risk. Spoiler alert: we found in our example that it surely does.
Our study assumed a portfolio with a developed-market benchmark of the STOXX Global 1800 ex-USA index that was allowed to invest up to 10% in emerging markets. We created an optimized portfolio that aimed to maximize exposure to the WW4 Profitability factor with a 3% active risk target. The investment universe was the benchmark, plus a core emerging markets universe. There was a maximum active country weight of +/- 3% and an upper limit of 10% emerging markets allocation. There were no other constraints (e.g. turnover). We created two portfolios, one using WW4 and the other using LM4. The portfolios were rebalanced monthly from January 2000 through December 2020.
Same alpha but two risk models
Our main goal for this test was to determine which of the two risk models provided more accurate risk forecasts. But given that the alpha input was the same in both, we also wanted to ensure we did not sacrifice performance. Summary results (Figure 1) show that while the LM strategy produced a slightly lower realized active return, it did so with lower risk that was just about spot-on with the 3% goal, and therefore produced an identical information ratio. It did have slightly higher turnover and 10 fewer names on average, but the results overall were quite acceptable.
Figure 1. Backtest Summary Results
Developed and emerging markets factor behavior does differ
Diving a little deeper into the data, we find more evidence that by recognizing the differences in factor and industry risks across the two regions in the investment universe, the linked model bettered WW4 in a number of ways.
First, we note that WW4 assumes that Profitability in developed markets ex-US and emerging markets will exhibit the same behavior. But we found this not to be the case. In fact, the monthly returns have a relatively low correlation of 0.22 (Figure 2). Since 1996, the average annualized return for Profitability was 1.90% in DMxUS4, 2.52% in EM4 and 2.35% in WW4. In addition, although they have converged recently, the level of volatility in each region was quite different over time, with Profitability’s volatility in emerging markets mostly higher than in DMxUS, and WW4’s volatility almost always lower. It is clear that the lack of correlation suggests risks can offset each other, and that recognizing the higher volatility in EM should lead to a better overall risk forecast by dampening exposure where necessary.
Figure 2. DMxUS and EM Returns and Volatility
The linked model also recognizes that other industry and factor bets can offset each other. Although aggregate county risk is a relatively small proportion of the overall risk of the portfolios (Figure 3), emerging markets was actually a diversifying bet in the portfolio optimized with LM4 (Figure 4). That diversification benefit was not available in WW4, where the emerging markets allocation was almost always additive to overall risk. Again, if bets can offset each other because they have low, zero or negative correlations, the overall estimation of risk should be more accurate.
Figure 3. Aggregate Country Contribution to Active Risk*
*Covariance is distributed across all factors.
Figure 4. Total Regional Contribution to Active Risk*
*Covariance is distributed across all factors.
Cut to the chase: LM risk forecasts were more accurate
There are many other risk metrics we could examine, but since our goal is to illustrate how using the linked model can provide better risk forecasts, we will cut to the chase. As we noted at the beginning, over our full test period, realized risk from the linked model was spot on with the desired level, whereas when we used WW4 it was about 10% higher. Most managers don’t have 20 years to prove their mettle, however.
On a rolling 12-month basis, realized tracking error from the LM4 portfolio was almost always lower than it was for the WW4 portfolio. Of course, it is not necessarily a positive if realized tracking error is too low, but the gap between the two was highest when realized risk was well above the 3% target; the figures were much closer to each other when they were below 3%. The same held true on a 36-month basis, especially in the last five years of the test, when realized tracking error was trending up.
Our last comparison is of maximum 12-month drawdown. Over the course of our backtest there were five periods of drawdowns more than 1%. At the beginning of the test period, both portfolios fell the same amount. But over the next four the drawdowns were always bigger for the portfolio constructed with WW4. Although this is too few observations on which to base any firm conclusions, at least they are all in the right direction: the linked model would have saved about 1% of pain in each.
Figure 5. Rolling Realized Tracking Error
In this post we have presented another case study highlighting the benefits of a risk model that links regions over a single global model. Our study showed that using a linked model on a portfolio with the ability to invest outside of developed markets, but maintaining the developed markets benchmark, clearly benefitted the risk forecast accuracy. And this occurred without a major sacrifice to return and no deterioration in information ratio. The linked model recognized the diversification benefits from low-correlated returns, along with the need to rein in bets on a region where the factor’s forecast volatility was higher. This test used the Profitability factor, although we ran the same test using Value and Momentum and came to the same conclusion.
 In other words, it had a negative contribution to risk.