I have come across two ways of measuring VaR for Fixed Income instruments thus far:
Express the volatility in of basis points and the position in terms of sensitivity to a 1 basis point movement in yields and then multiply it by the desired largest possible movement (95% or 99%); This method is described on page 17 of this document.
Map the cash flows of an instrument (a coupon bond) into buckets, get the zero rates (interpolate if needed), find PV01, volatility of these zero rates, get the correlation matrix for the zero rates, find the total variance and calculate the VaR.
The first one seems relatively simple. However, what happens if there are many bonds in the portfolio? Is it OK to find individual VaRs for each bond and then simply sum them up?
On the other hand, the second approach does deal with covariance of rates in different time horizons. But there might be a small error due to bucket specification. Theoretically, we could get an infinite number of buckets. But this is obviously very complicated.