Overwatch

Inspiration

Polymarket bets on the news. We made the city itself tradeable.

London's next 5 minutes become a live market, settled by REAL data from TFL, not opinion. AI traders read the city's live feeds; the market learns which to trust. The result is new data: a live price on what a city does next.

Marshall Wace built TOPS to learn which brokers to trust. We built it to learn which data reads a city best.

What it does

New markets every 5 minutes: busiest hub vs its usual day, Tube-wide busyness, the London VIX, hourly headlines.
Settled by sensors: only on readings TfL stamped after betting closed. Nobody knows the answer early. No reading, full refund.
7 AI traders, each reading one feed: crowding, cameras (YOLO), Tube status, rain, Heathrow planes, the crowd, our simulation.
TOPS on a city: reality scores every call, and only proven traders move the price. No record, no say.
The Exchange: same markets, no house, an order book against 2,196 AI traders (market makers, funds, retail).

How we built it

16 live data sources, crowding polled every 30 s, YOLO counting people on 13 cameras.
Prices: 50 possible futures per market, reweighted by every new reading, then an LMSR market maker.
AI traders: Kelly-sized bets; each one's skill is a Bayesian estimate, and the market only leans on them once their record beats our model by more than its noise.
Fair by construction: bets fill only after our next data poll, so nobody can trade on data we haven't seen.
The Exchange: real order books. On markets with a known answer, prices end 52% closer to the truth than our own model.
Agent simulation: 50 worlds on 1,017 real streets. It trades as one AI; it stays out of the price until it earns its place.

Challenges we ran into

Markets were too easy. 286 of 309 opened at 90%+. We set lines from the data so every market opens near 50/50.
Bots beat the house, so we let them. They found 3 leaks; we closed 2 and a timing edge (+124% → +8%).
Our forecast tied "nothing changes". New rule: every model must beat the old one and a coin flip out of sample, or it doesn't ship.

Accomplishments that we're proud of

Live on real London data, around the clock.
275 markets replayed, 0 voids.
The market tells good data from bad: it caught our own simulation trader losing to the model over 78 markets (99.9% confidence) and gives it no say.

What we learned

A good market is one you only win by reading the data.
Beat the baseline before claiming anything.

What's next for OverWatch

Move live trading to the Exchange once it prices London better on recorded data.
Real-money licensing, business hedging, the prices as alternative data.
Every city: all it needs is a street graph and live feeds.