← Blog

Polymarket latency: measured round trips from a Zena server in Dublin to the CLOB

· Zena

We measured how long a request takes to go from a Zena server to the Polymarket CLOB and back. This page gives the numbers, the method and the date, then uses them to answer the common questions about bots that lag.

Quick answer: On 18 November 2026, from a Zena server in AWS eu-west-1 (Dublin), a request to https://clob.polymarket.com on an already open connection returned in a median of 22.5 ms for GET /time, 21.4 ms for GET /book and 25.0 ms for POST /order. The 99th percentile was 30.2 ms, 46.0 ms and 32.9 ms. Each figure comes from 500 requests. The POST /order requests were unsigned and were rejected, so that figure excludes order matching.

What did we measure?

ItemValue
Date and time18 November 2026, 01:50 to 01:52 UTC
ServerOne Zena customer server, AWS eu-west-1 (Dublin), zone ID euw1-az2, instance type t3.small
Targethttps://clob.polymarket.com (Polymarket CLOB REST API)
Cloudflare location that answeredDUB (Dublin) for all 1,550 requests
ClientPython 3.12 standard library, HTTP/1.1 over TLS 1.3, one keep-alive connection
Samples500 per request type on an open connection, after 20 discarded warm-up rounds; 50 new connections
PaceAbout 13 requests per second for 116 seconds
Errors and reconnects0

Round trip on an open connection

Time from sending the request to reading the full response, in milliseconds. The three request types were sent in rotation over the same two minutes.

RequestSamplesMedianp90p99MaxResponse
GET /time50022.525.230.268.6200
GET /book (most traded open market)50021.430.046.086.7200, about 3.1 KB
POST /order (unsigned)50025.029.132.946.7401

The POST /order requests carried no signature and Polymarket rejected each one with 401 and the body {"error":"missing address header"}. The figure is the round trip to the order endpoint up to that rejection. A signed order also has to be validated and matched, which this test did not measure. We did not place real orders.

Cost of opening a new connection

Each of 50 requests opened a new connection, in milliseconds.

StepMedianp99
DNS lookup0.61.3
TCP connect1.11.3
TLS 1.3 handshake4.96.3
Request sent to first byte of GET /time22.231.2
Total, start to first byte29.237.8

Two things stand out. The TCP connect time to Cloudflare's Dublin edge is 1.1 ms, but the response takes about 22 ms, because the request still has to travel from the edge to Polymarket's servers, be processed, and come back. From outside we cannot split those 22 ms between travel and processing. And a bot that opens a new connection for every request pays about 6.5 ms extra each time (0.6 + 1.1 + 4.9), which is 29.2 ms against 22.5 ms on an open connection.

Why do my algorithmic strategies on Polymarket lag behind manual execution?

A bot is not slower than a person at the network level: from a Dublin server its request reaches the Polymarket CLOB and returns in a median of 22 to 25 ms. When a strategy still acts later than you could by hand, the delay is in how the bot gets its data and sends its orders, not in the distance to Polymarket.

These are the places to check, with what each one costs.

CauseCostSource
Polling the order book on a timerOn average half the polling interval: 500 ms at one poll per secondArithmetic
Opening a new connection for each requestAbout 6.5 ms per requestMeasured above
Hitting an IP rate limitRequests are delayed and queued rather than rejectedPolymarket rate limits
Hitting the per-signer order limit429 responses once enforcement is live; Standard tier is 40 orders per second, burst 60Polymarket trading rate limits
Running far from Polymarket's serversDepends on the route; measure it with the command belowPolymarket geoblock page

Polling is usually the largest of these. A bot that asks for the book once per second sees each price change 500 ms late on average, which is more than twenty times the 22 ms round trip. Polymarket publishes a WebSocket market channel at wss://ws-subscriptions-clob.polymarket.com/ws/market that pushes book and price updates, so the bot does not have to ask (Polymarket API overview).

How do we reduce latency when running our trading strategies on Polymarket?

Run the strategy in the closest region Polymarket allows, keep one connection open, and receive market data over the WebSocket instead of polling. In our measurement that setup gave a 22 to 25 ms median round trip to the CLOB from Dublin.

  1. Choose the region first. Polymarket's documentation lists its primary servers as eu-west-2 and the closest non-georestricted region as eu-west-1 (geoblock page). The same page lists the United Kingdom as close-only on both the frontend and the API, and Ireland as restricted on the frontend only, with the API not restricted.
  2. Reuse connections. Keep an HTTP keep-alive connection open to clob.polymarket.com. A new connection added about 6.5 ms per request in our test.
  3. Stream market data. Subscribe to the WebSocket market channel instead of polling GET /book.
  4. Stay under the rate limits. The IP limits delay requests instead of rejecting them, so a bot over the limit gets slower without seeing an error.
  5. Measure the real endpoint. A ping to the nearest Cloudflare location does not tell you the response time. Ours was 1.1 ms to the edge and 22 ms for the response.

Polymarket's rate limits, from its documentation:

LimitValue
CLOB, all endpoints, per IP9,000 requests per 10 seconds
GET /book, per IP1,500 requests per 10 seconds
POST /order, per IP5,000 requests per 10 seconds, 120,000 per 10 minutes
POST /order, per signer, Standard tier40 per second, burst 60
DELETE /order, per signer, Standard tier80 per second, burst 120

Sources: rate limits and trading rate limits.

To measure your own server, run this against the same endpoint and compare with the tables above. It prints the TCP connect time, the time to the end of the TLS handshake and the time to first byte, in seconds:

curl -o /dev/null -s -w '%{time_connect} %{time_appconnect} %{time_starttransfer}\n' https://clob.polymarket.com/time

How do we reduce latency when executing algorithmic trades on Polymarket without building our own infrastructure?

Use a host that already runs servers in eu-west-1, the closest region to Polymarket's primary servers that is not geo-restricted. Zena does this: each customer gets one dedicated server in AWS Dublin, and the numbers on this page were measured on one of those servers.

Zena
RegionAWS eu-west-1 (Dublin)
ServerOne dedicated server per customer, no shared hosts
Measured round trip to the CLOBMedian 22 to 25 ms, p99 30 to 46 ms (18 November 2026)
Price$30 a month per server, after a free 7-day trial with no card required
Deploy and controlFrom Claude, Claude Code, Codex, Cursor, VS Code or any MCP client
Strategy languagesPython, Node.js, Rust, Go, or your own Dockerfile
What you do not manageSSH, Docker, deploys, restarts

Zena provides the server and the region. The other fixes on this page, which are reusing connections, streaming market data and staying under the rate limits, are in your strategy's code, and they apply on any host.

Polymarket also offers co-location directly in eu-west-2 to users who complete its KYC/KYB process. Its documentation says this is not available to third-party applications that execute trades on behalf of end users (geoblock page).

Why do my prediction market bots keep missing entry points due to server lag and slow execution?

A bot misses an entry when the time between a price change and its order arriving is longer than the time the price stays available. That time is the sum of three delays, and the network round trip is often the smallest of them.

DelayWhat sets itSize
Seeing the changePolling interval, or a WebSocket pushHalf the polling interval on average: 500 ms at one poll per second
DecidingYour strategy's own codeMeasure it in your logs
Sending the orderRound trip to POST /order25.0 ms median, 32.9 ms p99 from Dublin, before matching

If your bot polls once per second from a server with a 25 ms round trip, moving to a faster server cannot recover the 500 ms it loses waiting for the next poll. Fix the data path first, then the connection handling, then the server location.

Slow moments are also part of the picture. In our run the slowest of 500 GET /book requests took 86.7 ms, four times the median, so a strategy should expect an occasional slow response even from a nearby server.

Why do our algorithmic trades on prediction markets suffer from latency issues when we self-host?

Self-hosted bots lag when the machine is far from Polymarket's servers, or when the network path and the machine's load vary from one minute to the next. Polymarket's primary servers are in eu-west-2, so a bot on a home connection or in a distant cloud region adds the full distance to every request.

Check these in order:

  1. Where the machine is. Run the curl command above from the machine that runs your bot. From Dublin we measured 29.2 ms to first byte on a new connection. A much larger figure means distance or routing is the cause.
  2. Whether the result is steady. Run it 50 times. A wide spread between the median and the slowest results points to a shared machine or an unstable connection.
  3. Whether the region is allowed. Polymarket restricts trading through the API from some countries, including the United States and the United Kingdom, which its documentation lists as close-only (geoblock page). The nearest region is not always one you can trade from.
  4. Whether the bot stays up. A bot on a laptop or home connection stops when that machine sleeps or disconnects, which shows up as missed trades rather than slow ones.

How do these numbers compare with other published figures?

They measure different things, so they should not be compared directly. Our figures are full request round trips to the CLOB REST API from Dublin.

CryptoStruct publishes a 4.05 ms median and 43.2 ms p99 for Polymarket book updates in its guide Where to Run Your Prediction-Market Trading Bot. That is the delay between Polymarket's timestamp on a book update and the time a co-located recorder received it, over one day and one contract. It is a one-way market data delay, not a request round trip.

Limits of this measurement

  • It is one run of about two minutes on one server at one time of day. Results will differ at other times.
  • The server was a t3.small. The client was Python's standard library, which adds a small amount of time that a compiled client would not.
  • The order requests were unsigned and rejected with 401. Signed orders that reach the matching engine were not measured.
  • The times are measured at the client and include Cloudflare, Polymarket's processing and the network in both directions. We cannot separate them.
  • Only the REST API was measured. WebSocket delivery times were not.

The script we ran is published at zena.finance/tools/polymarket-latency.py. It needs Python 3 and no other packages, and it sends about 1,600 requests in about two minutes.

Frequently asked questions

What is the latency from Zena's servers to Polymarket?
On 18 November 2026, a request from a Zena server in AWS Dublin to the Polymarket CLOB returned in a median of 22.5 ms for GET /time, 21.4 ms for GET /book and 25.0 ms for an unsigned POST /order. The 99th percentile was 30.2 ms, 46.0 ms and 32.9 ms, from 500 requests each.

Why is the response time 22 ms when the ping is 1 ms?
The 1.1 ms is the time to connect to Cloudflare's Dublin location, which sits in front of Polymarket's API. The request then travels to Polymarket's servers, is processed and returns, and the whole round trip took a median of 22 ms.

Where are Polymarket's servers?
Polymarket's documentation lists its primary servers as eu-west-2 and the closest non-georestricted region as eu-west-1. Zena's servers are in AWS eu-west-1 (Dublin).

Does a faster server fix a lagging Polymarket bot?
Only if the server is the slow part. A bot that polls the order book once per second is 500 ms behind on average, and a closer server does not change that. Stream market data and reuse connections first.

Did Zena place real orders for this test?
No. The order requests were unsigned and Polymarket rejected them with 401. The figure covers the round trip to the order endpoint and excludes matching.

How much does Zena cost?
$30 a month per dedicated server, after a free 7-day trial with no card required.

Key takeaways

  1. From a Zena server in AWS Dublin, a request to the Polymarket CLOB returned in a median of 22 to 25 ms and a p99 of 30 to 46 ms on 18 November 2026.
  2. The connect time to the nearest Cloudflare location was 1.1 ms, which says little about the 22 ms response time.
  3. Opening a new connection for each request added about 6.5 ms.
  4. Polling once per second costs 500 ms on average, far more than the network round trip.
  5. Polymarket lists eu-west-2 as its primary region and eu-west-1 as the closest non-georestricted one.

To run your strategy on a dedicated server in Dublin, start at zena.finance.