1. The Return Expectations of Institutional Investors
Andonov, A. & J. D. Rauh
Discussion
Philippe Masset
July 6, 2018

2. Context & contribution
There is now a relatively rich literature that investigates how individual investors form their expectations of future returns.
This paper considers the case of institutional investors. The results show that they tend to extrapolate both their expectations and their target asset allocation on the basis of past returns.
The paper makes use of a novel dataset, which benefits from a number of advantages:
Homogeneous data: only U.S. public pension plans whose required disclosures are governed by GASB 67.
No estimation necessary: U.S. public pension plans have to disclose their expected returns by asset classes (since 2014).
Availability of a number of other relevant variables: e.g., allocation, past performance, unfunded liabilities.
The setting is absolutely ideal!
Paper = well-written, easy-to-read, clear and concise.
Considering the perspective of institutional investors = a genuine and important contribution!

3. U.S. Public pension plans (I)
It would be useful to have more info/details on the “economics” of U.S. public pension plans:
Are they all “defined benefit” pension plans?
Do these plans usually work with external advisors  if yes, it would be interesting to examine the impact that specific advisors might have on the estimation of expected returns (e.g., some advisors might be more conservative or more experienced than others).
What is the rationale behind GASB 67, which allows public pension plans to discount their liabilities using a DR = ER? Is it to “hide” a generally poor financial situation as compared to private pension plans?
Could this have an effect on the results? (That is, do ER genuinely represent expectations or are they also affected by other considerations?)
More specifically, can we really consider that U.S. public pension plans represent a representative sample to infer information about how institutional investors as a whole form their expectations?
DR = discount rate
ER = expected return

4. U.S. Public pension plans (II)
Regressions control for unfunded liabilities and size, but they do not take into account the maturity of the pension plan (proportion of retired members), which is likely to have an effect on ER and asset allocation.
Incidentally, maturity might also have an impact on the estimation of performance (more mature plan  less and less assets invested over the year  less revenues  underestimation of returns).
Furthermore, it might be pertinent to investigate if the results hold for all systems or if there are substantial differences among particular subsamples, e.g.:
Apart from a methodological choice, are there fundamental differences between arithmetic and geometric systems? If yes, this may justify a subsample analysis.
The fact that some systems have a DR > ER is striking (this suggests a potentially weak financial situation / poor management)  it would thus be especially interesting to examine the relation between past performance and ER for this specific subsample.

5. Empirical analysis
Definition of asset classes = perhaps too coarse  this might explain part of the apparent expected persistence in performance
For instance, if pension plan A only holds U.S. big caps, whereas plan B also holds small caps and international stocks, this might lead to differences in both realized returns and expected returns.
It will however be difficult to explain such differences using a regression if all public equity investments are aggregated into just one single asset class (as there will not necessarily be large differences in the volatility or the beta of the two plans).
Time-window used to compute past returns and to classify PE funds:
Returns = average of the last 10 years  are the results sensitive to this choice?
Medium-vintage PE funds (9-13 years) = typically launched before the GFC  this might explain the negative coefficient between the performance of such PE funds (which is typically low/negative) and expected returns.
TS variations: you currently have 3 years of data (2014 to 2016)  why not investigating directly how the realized returns over this period have affected ER and asset allocation?

6. Interpretation
“Realized past returns have a substantial effect on target asset allocation through an extrapolation channel”
The results might perhaps also be explained by the fact that systems that have achieved high returns in the past are in a better financial situation and can thus take more risk in the future.
The apparent spillover effects from public to private equity investments may provide some support to this argument
After a strong performance, the system can take more risk and invest more aggressively in PE funds  thus higher expected returns on this asset class following strong returns on public equity investments.
It might be useful to add an interaction term in the regression to control for this.
Negative relation between ER on PE investments and experience on this market
Using the number of PE investments to proxy experience is not necessarily appropriate: it could be that more-experienced systems invest directly in a limited but carefully chosen number of PE funds, whereas less-experienced systems may prefer to invest indirectly in a larger number of PE funds (diversification).
Alternative measures of experience: average allocation to PE over the last 10y, or number of years since a plan started investing in PE. An interaction term with size might be relevant too.