The source and load are supposed to be 50 ohms, and their voltages will drop by 50% when matched. That's simply power transfer at work.
Insertion loss is normalized to this. In a simulation, you'll usually start with a 2V source (or 2V/50 ohms current source, for a Norton equivalent), which when loaded with 50 ohms gives 1V and 20*log(1/1) = 0dB insertion loss, of course. If you get less than 1V output from your filter (peak in the passband), that's loss. (You probably won't measure any with ideal simulated components.)
So topology, those values are plausible but you'll need quite high Q, which isn't so plausible. Narrow filter you want a coupled resonator topology. Example:
https://www.jrmagnetics.com/rf/doubtune/doubccl_c.phpOf course this is only a 2nd order filter, not 3rd.
Basic design, consider: 1kHz width, 77kHz center, 77/1 = Q required. The actual loaded Q of each resonator may be higher or lower than this, but they will all average to this (geometric average). The loaded Qs are spread apart a bit to get the desired filter profile (Butterworth in this case). The component Q must be many times higher than this -- if you use an ideal capacitor, and an inductor of Q = 77, you have 50% loss on each resonator, you'll actually get a loaded Q of 38, and you get a very sloppy bandwidth of more than 2kHz. Probably a lot worse than that because there's three of them.
The impedance of the resonators doesn't matter much, because you'll be using an impedance matching network to go from 50 ohms to the resonator. The above uses a capacitive divider. Tapped or coupled inductors can be used, or any other divider sort of network you can put together without adding more resonant modes.
The impedance of a resonator doesn't matter too much anyway, because the number of turns and thickness of wire is variable, for an inductor of a given size. Nominal impedances (say 20 to 2000 ohms) are easier to implement with real materials (because of resistivity of copper, parasitic capacitance and so on).
But really, for such a narrow bandwidth, you may find it's better to use a wide filter that's easy to build, and use another method to narrow the bandwidth further. Example, say you use a 10kHz bandwidth (which can be implemented with Q ~ 30 commercial parts just fine), then detect it, or hetrodyne it down to baseband, and use an active filter (Rs and Cs and opamps). Or do it DSP with a microcontroller doing equivalent-time sampling, which might not be doable on an ATmega based Arduino, but pretty much anything ARM based will do.
Or sticking with passive filters, you're probably not going to find a mechanical filter or resonator at quite that frequency (unless it's a standard thing that I don't know about -- in which case, bonanza!), but upconverting it to 455kHz to use AM BCB filters, or the low MHz for quartz crystals*, would be another option.
*Or 32.768kHz, but you'd probably never get the bandwidth wide enough? The impedance is really high, too.
Tim