I'm calculating travel times by simulating a rider starting from the selected stop at a specific time (for example, "weekday late night" is 3 AM, and "rush hour" is 8 AM) and "waiting" for he next scheduled train. That's why, when you select most stops during late night, the nearest stops are pretty far (~ 10 minutes) away — because that's a typical time you'd wait for a train at that time of night. It's definitely not a perfect prediction.
And yep — I really need to put some work into the mobile experience. I'm surprised it works at all!
Going from some stations to stations on other lines, it takes more time to get to the transfer station than to several stops farther on the second line. That seems off.
Choose Rockefeller Center then look at the purple line just below it for example.
To address that issue, a better calculation might be to compute possible wait times based on distribution of likely arrival times at the station and return some weighted average.
The issue with using average waiting times (as opposed to picking a particular time and computing actual waiting times) is stations with multiple trains — the N, R and Q might each stop at Union Square once every 10 minutes, but if you're happy boarding any of those, then your average wait time isn't 5 minutes — it's much less.
What I might try is sampling a couple random start times (e.g. 8:00 AM, 8:03 AM, 8:22 AM) and averaging the predictions of all those.
And yep — I really need to put some work into the mobile experience. I'm surprised it works at all!