Moscow Transport Hackathon · 2026

See the delay
15 minutes
before it happens

The system listens to bus telematics over NDTP, matches every ping against the timetable and warns the dispatcher about a delay 10–15 minutes before the stop — with the cause and a recommendation for what to do.

Open the dashboard Docs Swagger

on time at risk will be late 13 real routes from the dataset

—

buses sending NDTP right now

—

telemetry packets received

—

active alerts

—

forecasts inside the 10–15 min window

—

ms per forecast (p50)

Numbers come straight from api.mowtransit.ru/health

01 · Horizon

Every forecast looks at the (T+10, T+15] min window

Every 5 seconds, for every bus, we take the first scheduled stop that falls inside the window — and ML forecasts the delay at that stop. As time moves, the target rolls forward and the forecast updates. No after-the-fact forecasts: an alert is only raised while the stop is still 10–15 minutes away.

T = 08:00:00

forecast window

T+5+10+15+20 min

Forecast target—

Scheduled—

Lead time—

Predicted delay—

100%of forecasts on the live stream land in the window — the service checks this itself and reports horizon_ok_share in /health

lead_minin every ML response: how many minutes before the scheduled stop the forecast was made

3 of 3verified alerts in a run on the real day turned out to be real delays (253–278 s)

02 · Incident card

Not just “it’ll be late”, but why and what to do

An alert is raised when the model puts the probability of being more than 2 minutes late at ≥ 0.7. The card shows the forecast with a range, the route segment, each cause’s contribution in seconds and a recommendation for the dispatcher. When the bus arrives, the alert is checked against the actual time.

The forecast range is calibrated: the actual delay falls inside it 80% of the time
Causes come from approximate SHAP, turned into plain language
The recommendation works out the speed needed to get back on schedule

The card is rendered straight from the live system, which talks to Moscow dispatchers — so its text is in Russian.

sample

03 · How it works

Stream → forecast → dashboard, end to end

Three independent services in Docker Compose. The backend decodes binary NDTP itself (no off-the-shelf parser); ML is a separate service that can be retrained and reloaded without stopping the backend.

📡

NDTP terminals

The organisers’ emulator replaying real tracks. Binary TCP to ndtp.mowtransit.ru:9201

13 buses live

⚙️

Backend · ingest

Our own NPL/NPH codec, CRC-16, handshake, G6CellNav00 cells. asyncio + uvloop

0.3 ms p50

🧭

Backend · state

Timetable matching, GPS-detected arrivals, current deviation, segment speed, dwell time

±3 s median

🧠

ML · CatBoost ×5

An ensemble on 56 features + range and probability models, causes, what-if

44 ms /predict

🗺️

Dashboard

MapLibre, risk traffic light, alert feed, what-if. WebSocket, updates every second

app.mowtransit.ru

All history — telemetry, arrivals, every forecast and alert — is written to Postgres in the background (write-behind, COPY), so the database can never slow down the live path.

Python 3.12CatBoostFastAPIasyncio + uvloopONNX PostgresDocker ComposenginxMapLibreWebSocketSphinxSwagger

04 · Accuracy

Off by 40 seconds where the baseline is off by 93

MAE of the delay forecast on a held-out set. The hackathon metric hits its ceiling: score = 1.0, both locally and on the platform.

Predict “no delay” (0 s)

103 s

Organisers’ baseline (cur_dev_s)

93 s

Our CatBoost ensemble

40 s

1.0

hackathon score
6 of 6 points for accuracy

38.8 s

Shaped like validate

5 folds in 30-minute blocks: nearby moments of the same buses. Baseline: 88.4 s.

78.9 s

Honestly: an unseen bus

The whole bus is held out — the model has never seen it. Still beats the baseline (88.4 s).

0.97

AUC for “will it be late”

The late probability is calibrated: predicted ~0.89 → actually 0.95. It drives the risk traffic light.

Online = batch: the live service matches submission.csv on 151 of 151 points, maximum difference 0.00 s. The same function computes the features in both.

05 · Reliability