attention flow

A live view of where traders' attention is going on Polymarket, and what it sets off.

It follows the busiest markets trade by trade and keeps two models fitted on the last two days:

• Most likely to move: the markets the model gives the best odds of a 2¢ price move in the next 10 minutes, compared with a normal moment, and which related markets are driving it. This held up out of sample in the last four months of a year of real trades, but not earlier in the year (see Real data).
• What sets off what: which markets' trading kicks off trading in other markets, and how fast.

Try it

You need Go 1.24 or newer. No API keys, no accounts: everything comes from Polymarket's public endpoints.

git clone https://github.com/gauri-sharmaa/attention-flow
cd attention-flow
go run ./cmd/attnflow serve

Open http://localhost:8080. It spends a minute or two loading recent trades and fitting, then goes live. -markets 60 -hours 12 warms up faster, and -addr :9000 changes the port.

Watch the 2-minute demo: how it works, what each panel means, and what the research found.

Real data

A year of every trade on 1,205 Polymarket markets, Oct 2025 to Oct 2026: 24 million trades, including markets that resolved during the year. Every test is walk-forward: fit on one 30-day period, scored on the next, which the model never saw. That gives 11 test months, or 12 periods for the tests that don't fit a model. Each test has a placebo that shifts streams by whole days, keeping their daily rhythm and destroying real timing.

Question	Method	Test months that pass	Answer
Does trading activity predict price moves?	Hawkes, trades → 2¢ moves	4 of 11 (Jun–Oct 2026); beats the placebo in 10 of 11	Recently, yes; earlier in the year, no
Does trading in one market set off related markets?	Hawkes, trade times	6 of 11	Some months
Do some prices move before related prices?	Hoffmann–Rosenbaum–Yoshida	12 of 12 (27–113 linked pairs a month, 1–4 by chance)	Linked, but the leader flips
Is the order-flow signal profitable?	paper trading, 1¢ per side	0 of 11 (−$219k on 20,118 trades)	No
Does news / Reddit / HN buzz lead trading?	event study, 1,743 surges (90 days)	–	No

A month passes when the held-out gain (log-likelihood per event, in nats) beats both zero and the placebo.

• Activity → price moves: gains of +0.10 to +0.20 per move from June to October 2026, against placebos of −0.13 to −0.54. From November to June the gain was negative every month (−0.01 to −0.19), though still above the placebo in six of those seven months. So the timing carries information all year, but only recently did it beat a model of price moves alone. The 90-day run covered only Jul–Oct, which is why it looked like a clean yes.
• Spreading between markets: clearly positive in Nov–Dec 2025 and Jul–Oct 2026 (+0.03 to +0.12), flat or negative in between.
• Prices: linked markets move together every month, but the leader repeats from one month to the next only 47% of the time, a coin flip. The 61 pairs with a lead that lasts are almost all the same question at different dates or strikes, where the busier contract reprices first.
• Trading: order flow has a little information before costs: +$13.6k over the year, positive in 10 of 11 months. A 1¢ cost per side turns that into −$219k.
• Buzz: around news and Reddit surges, trading in the related markets stays flat before and after.

Scorecards: docs/results (real-365d-* for the year, real-90d-* for the first run).

Bars blur timing, so the methods that work use raw event times:

• Hawkes processes (internal/hawkes): each trade raises the near-term rate of more trades in its own and related markets, and the fitted excitation matrix is the lead-lag graph. Same-second events never excite each other: a trade and the price move it causes are simultaneous, and blocking that cut the first cross-market estimate from +0.104 to +0.030. A shared hour-of-week background stops US daytime from passing for a link.
• Hoffmann–Rosenbaum–Yoshida lead-lag (internal/leadlag): covariance of price changes over overlapping observation intervals, scanned across lags, with a p-value from 99 shuffles.
• Event study (attnflow eventstudy): trading around each surge against the same clock time on other days.

Methods follow Bacry, Mastromatteo & Muzy, Hawkes processes in finance; Xu, Farajtabar & Zha, Learning Granger causality for Hawkes processes; and Hoffmann, Rosenbaum & Yoshida, Estimation of the lead-lag parameter from non-synchronous data.

Research commands

Go 1.24; the only dependency is a websocket library.

go build -o attnflow ./cmd/attnflow

./attnflow sim                                   # simulate 2 weeks with an answer key
./attnflow replay -truth data/sim/truth.json     # score everything
./attnflow serve                                 # live Polymarket dashboard at localhost:8080
./attnflow serve -sim                            # replay the simulator instead
./attnflow export -out site                      # static simulator dashboard for hosting

./attnflow polytrades -days 365 -n 400 -closed 1000 -bars=false -out data/pmt365   # a year of every trade, incl. resolved markets (hours; resumable)
./attnflow hawkes -dir data/pmt365 -folds 11      # does trading spread between markets? (walk-forward)
./attnflow hawkes -dir data/pmt365 -moves 0.02 -folds 11      # does trading predict price moves?
./attnflow leadlag -dir data/pmt365 -periods 12   # which prices move first
./attnflow backtest -dir data/pmt365 -periods 12  # paper-trade the order-flow signal
./attnflow eventstudy                            # does buzz lead trading?
./attnflow outside                               # news, Reddit, Hacker News mentions
./attnflow hawkes -extra data/pmt/outside-news.csv,data/pmt/outside-reddit.csv,data/pmt/outside-hn.csv

Real data (Wikipedia hourly pageviews for the topics in data/universe.txt):

./attnflow fetch -contact you@example.com -days 90
./attnflow replay -events data/wiki/events.csv -text data/wiki/text.json -bar 3600
./attnflow serve  -sim -events data/wiki/events.csv -text data/wiki/text.json -bar 3600 -label wikipedia

Any other feed (Kaito mindshare, Google Trends, attention-market prices) only needs to write the same ts,entity,value CSV.

Layout

cmd/attnflow       CLI
internal/engine    streaming inference: factors, RLS forecasts, lead-lag graph, shocks
internal/sim       simulator with a planted answer key
internal/replay    out-of-sample scoring, oracle ceiling
internal/semantic  TF-IDF candidate graph
internal/stats     online estimators (EW, robust median/MAD, RLS)
internal/ring      lock-free SPSC queue for ingest
internal/source    Wikipedia, Polymarket, Bluesky, GDELT, Reddit, Hacker News
internal/hawkes    multivariate Hawkes processes on raw event times
internal/leadlag   Hoffmann–Rosenbaum–Yoshida lead-lag for asynchronous prices
internal/server    SSE server + dashboard