In development · not yet launched
New York Penn Station · NJ Transit
You're not waiting
for a train. You're
waiting for a number.
Nostos is being built to work out your track before DepartureVision posts it — and to warn you when your train is at risk of being cancelled.
nostos (n.) — Greek. The homecoming.
The root of the word nostalgia.
| Dep | Destination | Line | Track |
|---|---|---|---|
| 17:51 | Trenton | NEC | — |
| 18:04 | Long Branch | NJCL | — |
| 18:14 | Dover AT RISK | MOBO | — |
| 18:22 | Bay Head | NJCL | — |
The problem
Everyone else answers
the smaller question.
Track prediction tells you where to stand. Useful. But it is not the thing that ruins your evening. The thing that ruins your evening is a train that never runs, announced eight minutes before it was due to leave.
A number, eventually
The track appears on the board a few minutes before departure. Three hundred people read it at the same moment and move at the same moment.
To know if you're going
Whether to leave the office now, grab dinner, or take the earlier train. Those decisions need twenty minutes of warning, not two.
The signs are public
Most Penn departures are inbound trains that arrived and turned around. When the inbound is late, the outbound is in trouble — and that is visible in the open data long before anyone announces it.
A cancellation, as it actually unfolds
- 16:58 — An inbound train loses time west of Newark. Nothing is announced. Nothing needs to be.
- 17:20 — It's now 38 minutes down. The equipment it was supposed to hand over is 38 minutes from where it should be.
- 17:41 — The turn window has collapsed. There is no longer time to unload, clean, and reverse before the 18:14 is due out.
- 18:06 — The 18:14 is cancelled. Reason given: equipment availability. Eight minutes of warning.
Nostos is being built to flag this at 17:20 — not by guessing, but by noticing the equipment can no longer arrive in time.
How it works
Three signals, in order
of how much they matter.
No black box, and less machine learning than you'd expect. Most of the work is bookkeeping done carefully.
Where the train physically is
NJ Transit publishes the track circuit each train occupies — the electrical block it is sitting on right now. Once a set enters certain circuits approaching Penn, the platform it is heading for is no longer a guess. At that point we are not predicting. We are reading.
Which train becomes which
Trains turn around. The 17:12 arrival becomes the 17:51 departure, usually off the same platform. We rebuild that equipment cycle from public data, so a late arrival tells us both the track and the risk for its outbound.
What normally happens
Your train has run five days a week for months. That history is a strong prior, and it is where every other tool in this space stops. For us it is the fallback when the first two have nothing to say.
One thing we do differently
Two trains cannot occupy one track. So rather than predicting each train on its own — which is how you end up confidently putting three trains on track 7 — we solve the whole next hour at once, as an assignment problem across every platform. Fewer confident contradictions, and a real reason behind each number.
Compared with what exists
There are other tools.
They answer different questions.
Track prediction at Penn is a crowded little market, and the tools in it are genuinely good at what they do. This is where Nostos is not the same.
The Nostos column describes what it is being built to do — it isn't available yet. The other columns describe tools you can use today.
| Official app | Other predictors | Nostos planned |
|
|---|---|---|---|
| Track before it's posted | No | Yes | Yes |
| Warning before a cancellation | No | No | Yes |
| Tells you without being opened | Alerts only | No | Push, widget, watch |
| Platform support | iOS + Android | Mostly iOS | iOS · Android · Web |
| Publishes its own accuracy | — | No | Every month |
| Price | Free | Subscription | Free |
On that last row but one
Any tool can claim it predicts your track. Almost none will tell you how often it was wrong. We publish a monthly accuracy figure, broken out by line and by how far ahead the call was made — including the months it gets worse. If we ever quietly stop publishing it, assume the worst and stop trusting us.
The business model
Free. No ads.
Nothing personal for sale.
This is built on a public agency's open data by someone who rides the Northeast Corridor. It should not cost commuters money to find out whether their train is running.
Charging for the answer
No subscription, no paywall on predictions, no premium tier that unlocks the thing you actually came for. Free is not a launch promotion.
Advertising
No ad network, no tracking scripts, no cookie banner, no interstitials while you're trying to read a track number in a crowd.
Selling anything personal about you
Your identity, device, location, and travel habits are never sold, shared, or used for targeting. We don't need an account to show you a board, and we'd rather not have one.
Taking money to change the answer
Nothing displayed is ever paid placement. A prediction is what the data says or it is worthless.
So what's the catch
You pay in corrections.
When a prediction is wrong, tap the button that says so. That's the whole arrangement. Every correction trains the thing that told you wrong, which means the people who use it most make it best. It costs you a second and it is worth more to us than five dollars a month.
To be precise about the word "data"
There are two different things people mean by that word, and we treat them very differently.
Anything about you — who you are, where you board, when you travel, what device you carry — is never sold, licensed, rented, or shared with anyone, and is never used to target you. We collect as little of it as we can get away with.
Anything about the trains — how often the 18:14 to Dover was cancelled last month, which lines run worst in winter — is computed from NJ Transit's public feed and describes the railroad, not its riders. We publish that openly, and we may one day license the analysis to newsrooms or researchers. That would be a business built on public information about a public service, and it will never involve anything personal.
If running costs ever outgrow a hobby budget, the first thing we'd try is a clearly-labelled sponsorship from a business near a station — a fixed placement, disclosed as an ad, never mixed into a prediction, and never chosen using anything we know about you. We'd announce it here before it happened.
Where this actually is
Not launched. Listening.
Being straight with you: there's no app yet — Nostos is still in development. NJ Transit keeps no public archive of track assignments, so the only way to build this is to watch the trains and learn the patterns first. That work is underway. We'll email you when there's something real to try.
Hear about it when it works
One email when there's something real to try. Nothing else, ever.
One email at launch, then nothing. We store only your address and never sell or share it — see our privacy notice.
Early riders on the Northeast Corridor and North Jersey Coast Line first — that's where the data is thickest.