Sizing risks in portfolio construction
The belief that investment strategies should be designed with a client’s objectives, constraints and time horizon in mind is at the heart of everything LGIM does.
Improving the range of possible outcomes for investors should, in our view, not simply be focused on delivering performance relative to a certain benchmark. Investors will hold a range of both short and long-term objectives, which could be wider than purely financial. So it is critical to understand the hurdles to an investor's ultimate success, not least economic and demographic challenges, the impact of costs, and risks such as the failure of a funding source.
Because sources of return need to be sized appropriately in light of these factors, we devote significant resource to engaging clients on the principles of portfolio construction, from establishing investment beliefs to diversifying rewarded risks and scenario testing.
DB: model portfolios
In this piece, I will focus on our work with UK defined benefit (DB) clients, showing below some illustrative strategies for a typical scheme.1
Our portfolios are driven by an emphasis on appropriate objectives and measures of success. We define success for pension schemes as the assets outlasting the liability cashflows and, in the event of failure, meeting as high a proportion of benefits as possible. This can represent a mindset shift for trustees who are used to thinking of risk in terms of more traditional measures.
Some graphs illustrate our approach. First, below is a fairly typical DB portfolio, very loosely based on the Purple Book from the Pension Protection Fund2, compared with our model portfolio. For these purposes, we assume rates and inflation hedge ratios of 62% of funded liabilities (i.e. assets) here.
Our model portfolio is built in two stages:
(a) We start by building a strategic portfolio, based on our long-term return estimates and modelling of risk. This includes modelling covenant (sponsor) risk; for this we have assumed a fairly typical initial sponsor rating of BB. Our approach involves testing how thousands of potential asset allocations perform in terms of the risk of ultimately failing to pay pensions.
(b) We then tilt the strategy, as illustrated below, by overweighting asset classes that appear to offer higher risk-adjusted returns over the medium term and underweighting those with worse apparent prospects. These tilts are based on LGIM's multi-asset views. The bar chart shows how the model portfolio differs from our long-term allocation. In general, our tilts vary with scheme-specific circumstances. For example, schemes with more aggressive strategic allocations have larger tilts.
What’s the impact?
Our process involves modelling the distribution of outcomes for members in terms of the proportion of pensions ultimately paid. We show below statistics based on this idea for long-term success, both on average and in a downside scenario. We also include some more traditional measures – namely the expected rate of return over gilts and funding level value-at-risk (VaR).
Our analysis suggests increased diversification3, using leveraged liability-driven investment to reduce ‘cash drag’ and increase hedge ratios, and ensuring an appropriate target level of return (that takes into account covenant risk) can improve the distribution of outcomes.
By introducing dynamic tilts, we estimate that our long-term success measures could be further enhanced from relative-value trades to reflect our macro views4. This framework can be used to analyse many types of investment strategy and scheme circumstances.
1. Buyout funding level 75%, 20 year duration
2. Note that this is based on the average of DB schemes so could overstate diversification.
3. Including at a more granular level than can be seen in the pie charts
4. The dynamic portfolio is expected to generate alpha of 0.5% pa relative to the strategic portfolio and on average (over time) is expected to have the same asset allocation as the strategic portfolio. To allow for active risk we allow for uncorrelated tracking error of 1% pa and, more significantly, allow for the parameter uncertainty in our mean alpha assumption. We assume ex-ante alpha that has a standard deviation of 0.5%; in particular we assume a c.16% chance that ex-ante alpha is negative.