I am wondering whether bond portfolio optimization within the mean-variance framework is meaningful. As is well-known, bonds of similar maturities tend to become highly correlated as maturity nears. Thus the traditional mean-variance practitioner's approach of estimating the covariance matrix from historical data would seem to be wishful thinking at best, and the mean-variance framework would seem to be on shaky ground for bond portfolios. On the other hand, I've heard that some reputable bond trading firms do in fact perform some form of mean-variance-like portfolio optimization. Martellini's latest book on "Fixed-Income Securities" also deals at great length on the subject of covariance matrix estimation for bond portfolios, with not even passing mention of the above problem.So my question is, what if any optimization is performed by bond portfolio managers in industry, and how prevalent are these approaches in practice? Many thanks for your feedback.

Does your management only care about mean and variance? On a "stand-alone" or "marginal contribution to portfolio" basis?Mean and variance of WHAT - economic value, accounting value, regulatory value?What about non-variance risk measures like maximum cashflow draw, probability of loss, etc?Mean and variance are academic measures that statistically nice to work with. Without further refinement, however, these measures don't really get at anything management really cares about.

Sorry, my original post appears to have been a bit ambiguous. I was referring to simply adopting the mean-variance framework for equity portfolios wholesale, with stocks now replaced by bonds---the usual objective of maximizing portfolio return subject to various constraints, e.g., shortselling, bounds on amount bought/sold for each asset, shortfall risk, minimum variance, etc. Or one could consider portfolio index tracking, e.g., minimizing tracking error. The great majority of the various extensions and refinements of Markowitz's original mean-variance framework have stocks in mind, and it seems to me that bonds present special problems that make it dangerous to apply mean-variance optimization without further consideration of these differences. For example, some literature I've seen on bond portfolio optimization implicitly recognizes that covariance matrix estimation for bond returns is fraught with dangers, and therefore eliminates covariances entirely in the problem formulation. With this clarification, I am looking forward to receiving your feedback.

For me it makes sense if you use bond indices instead of bonds

Ah, I see now what you are grappling with. This is a very interesting question. You have moved into a framework where yield curve dynamics (at a minimum) simultaneously impact return, volatility, and correlation. Add to this that interest rate models always seem to fall a bit short and interesting problems arise.(Interest rate models alway remind me of applied heat transfer problems. There are all of these complicated dynamics going on that we can't quite get ahold of, but we manage to muddle through most of the time. Unfortuantely, my heat transfer professor noted that I usually muddled more than necessary.)It's not clear to me, however, that bond indices solve the problem. I, myself, would be interested to see more discussion about this.

One thing to consider is decomposing bonds into the cashflows attached to fixed zero coupons along the yield curve. You then work with these instruments rather than bonds and use the price volatility implied by the invidual yields of the zero coupons. The problems with this approach are numerous - yields of zero coupons vary with both interest rates and the perceived credit rating. You have to maintain sets of zero coupons for a range of credit ratings and currencies. In principal it is quite straight forward though to maintain risk free yield curves (+implied price, vols and correlations) and all the spreads. But the key problem in my opinion is how to handle both credit rating volatility and interest rate volatility in a single factor mean-variance reduction problem. I'd be interest to know of any advances in this area if anyone knows.