Temperature is the largest factor. Things like a DS3231[1] do really well compared to a basic non-compensated oscillator. I have been running some long-term e

Temperature is the largest factor. Things like a DS3231[1] do really well compared to a basic non-compensated oscillator. I have been running some long-term experiments on a few that I have around and with some tuning got them to less than a second loss per year. But, they are super expensive compared to the basic ones (almost $5 each in quantity), so they aren't going to end up in your car where a 3 cent chip is possible to use instead. (I don't know what 5G / LTE chips cost these days, but if they're putting one in your car anyway, then they can probably get the time from that. But choose not to.) [1] https://www.analog.com/media/en/technical-documentation/data-sheets/DS3231.pdf https://www.analog.com/media/en/technical-documentation/data... Most interesting to me in all of my time experiments is looking at my clock frequency over time vs. the temperature. (NTP daemons aim to calculate your actual clock frequency; then they know how far off your internal time is from actual time.) You don't even need a temperature sensor, the clock rate is a perfect analogue.