Here’s an idea for some enterprising engineer (most likely
at Google or somewhere else with access to good traffic data) that I’m almost
certainly not the first to have thought of.
A good traffic prediction algorithm would let you specify a
time of day you need to arrive at a particular destination, a starting point,
and tell you when you need to leave. Google Now already does a crude
version of this. If you have flight details in your gmail account, it will sent
you an alert when you need to leave in order to get to the airport an hour
before your flight. But there’s a lot more cool stuff you could do with this.
For instance, it would be great to be able to take the
directions in Google Maps and specify a day of the week and time (or day of the
year) and see an estimate for how long the trip would take at that particular
point in time. Since google has oodles of historical traffic data, they’d be
able to get a pretty good estimate based just on historical traffic conditions.
Ideally, you’d be able to take the same route and plot out how the expected
length of journey varies with the starting time.
This would tell you what times of the day and night to
avoid, letting you figure out how to adjust your work schedule to avoid
traffic. It would also tell you about a fascinating quantity – the elasticity
of time arrived to time left. There are times of the day, such as peak hour,
where leaving 10 minutes later might cause you to arrive 15 minutes later (an
elasticity of 1.5, suggesting that wasting those minutes is very costly), or at
the back end when you can leave 10 minutes later and only arrive 8 minutes
later (making those minutes subsidised).
Notably, everything I’ve described (like Google Now in its
current form) only speaks of a point estimate of how long things will take,
presumably either the mean or median. In reality, there’s much more interesting
stuff you can do with the whole distribution.
For instance, lots of unexpected things happen with traffic –
accidents, weather, what have you. So for a trip that leaves at 8am on a
Monday, there’s actually a distribution of possible arrival times. For someone
who knows what a distribution actually means, it would be very useful to be
able to specify an acceptable percentage of the time that you would be late (or
more than X minutes late), and have the algorithm give you a time that you
needed to leave your house in order to get there on time with that probability.
If this were done, you could just subtract the number of
minutes you need to get ready each morning, and that’s when you need to set
your alarm.
Even more interesting would be to improve these predictions from
unconditional to conditional by making use of both current traffic and weather
conditions. The overall distribution of, say, Mondays in January, would give
you the unconditional distribution of the chances of arriving on time. But you could
definitely do better by generating conditional distributions that morning that
relied on the local weather conditions and the current traffic conditions
relative to the historical distribution. In other words, if you normally need
to leave home at 8am, the app could use the fact that traffic at 6:30am is heavier
than normal to estimate that you may need to wake up earlier than normal as
well.
Done properly, I’d gladly pay $20 for this kind of app. If
it really worked, I’d probably value it at much more than that, notwithstanding
that an irrational cheapskate instinct kicks in regarding the prospect of
paying more than a few bucks for an online app.
As with all Shylock ideas, should the app succeed I insist
on receiving either fat royalties or a free t-shirt that says ‘I came up with the
idea for [Traffick-ator] and all I got was this lousy t-shirt’. Medium please.
