The MSF set clock I built earlier this year was using the AVR’s 16MHz crystal for timekeeping. It was fine during the warmer months but recently as temperatures fell it started to loose accuracy. By the late evening it could be as much as 30 seconds out, so I decided to add a new reference clock with greater accuracy.
After evaluating a few different options I decided to use a CMOS 4060 14 stage ripple counter with a 32768Hz watch crystal. My experiments showed a very high degree of accuracy (better than 0.005%) at very low cost. The exact component values used for the resistors and capacitors did not seem to make much difference (±0.002%) either. Once installed in the clock case the environment was a bit harsher than my test bench due to EM and supply ripple, much of which is generated by the multiplexed display. Long time readers (ha) may recall that the display has to be turned off during MSF reception because of the noise it creates. Despite that the clock is now accurate to within 2 seconds per day, and since it is synchronised every night that is more than adequate.
For non-synchronised clocks a more accurate time source would be required for long term stability. Using the mains 50Hz frequency is a popular option. Realistically getting much better than 2 seconds/day is going to require some shielding and a well filtered power supply. Temperature does not seem to be as critical for watch crystals but even over room temperature ranges it will be an issue.