Short-Horizon Hedge Fund Allocation with Sparse Return Data
Summary
The document considers how to optimize a portfolio over a short horizon when hedge fund or other illiquid asset indices are available only at low frequency. It questions whether monthly observations provide enough data for a two-year estimate and rejects linear interpolation as a source of meaningful new observations. One proposed approach is to model a joint return distribution: combine non-normal marginal distributions informed by empirical data, Bayesian estimates, and mappings from liquid assets, then use Monte Carlo or a quasirandom sequence to optimize a chosen objective.
A second suggestion is hedge fund replication. If fund returns can be attributed to known risk-factor exposures, rolling-window attribution can estimate those exposures, while daily factor series provide a higher-frequency representation. These are alternatives rather than demonstrated solutions: the exchange supplies no empirical comparison, and both approaches rely on assumptions about distributions, factor relationships, and the stability of exposures.
Key ideas
- Interpolating sparse monthly returns creates synthetic observations rather than independent market data.
- A joint distribution can combine non-normal marginals with dependence modeled through a copula.
- Bayesian estimates and liquid-asset mappings can inform illiquid-asset return distributions.
- Monte Carlo or quasirandom sampling can support optimization over a modeled distribution.
- Factor-based replication can use rolling attribution and daily risk-factor histories.
Tags
Full text
# How do you handle short-term asset allocation with Hedge-Funds? # How do you handle short-term asset allocation with Hedge-Funds? Assuming I want to run an optimization over a short period, say 2 years, I would decide to take daily values in order to compute the efficient frontier of a portfolio. That works fine as long as I have classical assets with daily indices. For other asset classes, these indices might not exists. How would you then proceed? Do you use monthly values? If so, it looks like there won't be enough observations in the sample (only 24 for 2 years)... Would you then increase the period? I guess one could try to do a linear interpolation of the points during the months, but it would be useless as the resulting series would have more "generated" points than original ones. Is there another alternative? ## Answer by Brian B (score 3, accepted) https://quant.stackexchange.com/a/2463 I would not think of this in terms of modern portfolio theory at all, but rather include ideas that form the mathematical basis of MPT. In particular, construct a joint distribution for the various assets using some kind of copula, but with non-normal marginals derived from some combination of empirical returns, bayesian estimates, and "mapping" of illiquid asset return distributions to liquid ones. Once the joint distribution is specified, you can choose your favorite objective (which needn't be the same utility function as found in MPT) and optimize it over the joint distribution, likely using Monte Carlo and a quasirandom sequence. ## Answer by Tal Fishman (score 2) https://quant.stackexchange.com/a/2464 The alternative is hedge fund replication. Many hedge funds' returns may be broken down into (possibly time-varying) exposures to known risk factors. Andrew Lo has done much work on this topic. In principle, one may perform a returns attribution over a rolling window for some number of months, then use the daily time-series for the risk factors to fill in the gaps.
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.