> You are likely going to have to use some type of temp stabiliser to
> get the accuracy you require.  If you can spare the energy then you
> could use a heater (classic crystal oven principal) to hold the
> occilator at a temp of between 60 and 80 Deg.C with a simple
> controller.  Insulation would then reduce fluctuations and if would
> not spend a lot of time away from 60 C.

One approach I've hard of is to use a crystal together with an RC oscillator
and EEPROM table.  After the unit is assembled, a test fixture operates it
at different temperatures, measures the exact frequency of the crystal
oscillator, and measures the frequency ratio between the RC oscillator and
the crystal; these measurements are stored in EEPROM.

Later, in the field, the CPU can measure the RC oscillator versus the
crystal, ascertain the current temperature from this measurement, and then
use that to correct for any crystal frequency error.

For this approach to work correctly, it is imperative that the RC oscillator
and crystal oscillator be close enough together that temperature shifts will
affect them both similarly.  If not, a temperature shift that causes the RC
to run slow but not the crystal may have undue effects on the net frequency
output.