秋田预测市场 · Discord 审计

用户讨论交易延迟与Colocation对策略影响

2026-07-15 · Polymarket · 55 条相关讨论

Sampoman 2026-07-15 19:34:36

There's no way to get this sh*t running without co-located servers

There's no way to get this sh*t running without co-located servers

pUSD 2026-07-15 19:34:43

huh

huh

Sampoman 2026-07-15 19:34:47

You'll always lose in the latency game

You'll always lose in the latency game

Sampoman 2026-07-15 19:35:01

It is at least for 5 min btc windows

It is at least for 5 min btc windows

Sampoman 2026-07-15 19:35:21

When there's X price requests in front of the queue, you'll always lose

When there's X price requests in front of the queue, you'll always lose

Sampoman 2026-07-15 19:35:30

No matter how good your algo is

No matter how good your algo is

pUSD 2026-07-15 19:37:04

its a special case with polymarket i would say

its a special case with polymarket i would say

pUSD 2026-07-15 19:37:08

where it doesnt matter much

where it doesnt matter much

Sampoman 2026-07-15 19:37:44

Interesting. Been testing for at least 100 different algos with 5 min btc. Several of them work in sim but in reality you get front-run by other bots that are faster than you

Interesting. Been testing for at least 100 different algos with 5 min btc. Several of them work in sim but in reality you get front-run by other bots that are faster than you

Sampoman 2026-07-15 19:37:58

And I've been in Ireland so far

And I've been in Ireland so far

trex8791 2026-07-15 19:39:05

Me too. One critical thing I noticed is you cannot backtest maker strategies faithfully cause failed taker transactions are not recorded

Me too. One critical thing I noticed is you cannot backtest maker strategies faithfully cause failed taker transactions are not recorded

The Dude 2026-07-15 19:39:13

Maybe these algos sent their orders way before you, a time gap not explained by the Ireland London distance

Maybe these algos sent their orders way before you, a time gap not explained by the Ireland London distance

The Dude 2026-07-15 19:39:24

If it's 5-10 ms maybe colo explains it

If it's 5-10 ms maybe colo explains it

Sampoman 2026-07-15 19:39:34

It's really tight, that I can tell

It's really tight, that I can tell

The Dude 2026-07-15 19:39:39

If it's more you just have bad code/bad data

If it's more you just have bad code/bad data

The Dude 2026-07-15 19:40:01

It does not really exist anymore

It does not really exist anymore

Sampoman 2026-07-15 19:40:15

You can front-run BTC price changes by a few dozen ms, but so does everybody else

You can front-run BTC price changes by a few dozen ms, but so does everybody else

Sampoman 2026-07-15 19:40:28

And Ireland => London is 25ms

And Ireland => London is 25ms

Sampoman 2026-07-15 19:40:31

Edge gone, bang

Edge gone, bang

The Dude 2026-07-15 19:40:55

Dublin -> US is 45

Dublin -> US is 45

Sampoman 2026-07-15 19:41:05

AFIK Clob is in London?

AFIK Clob is in London?

pUSD 2026-07-15 19:41:06

isnt this ghost fill thing mostly old stuff

isnt this ghost fill thing mostly old stuff

The Dude 2026-07-15 19:41:25

But Dublin London is 5-10 ms maximum

But Dublin London is 5-10 ms maximum

Sampoman 2026-07-15 19:41:35

Still too much

Still too much

Sampoman 2026-07-15 19:42:00

I'm 99.99% sure there's bots running closer that drive the price

I'm 99.99% sure there's bots running closer that drive the price

The Dude 2026-07-15 19:42:03

Sure but I dont think it will make or break a strategy to have colo or not on PM

Sure but I dont think it will make or break a strategy to have colo or not on PM

Sampoman 2026-07-15 19:42:50

It's absolutely about that, because the price runs off usually by 3-4 cents by the time you react, so someone is sweeping the table

It's absolutely about that, because the price runs off usually by 3-4 cents by the time you react, so someone is sweeping the table

trex8791 2026-07-15 19:42:59

That's not what I mean. What I'm saying is many taker transactions fail to get matching orders and those do not get published. So you cannot backtest only against trade_tape and be faithful to reality

That's not what I mean. What I'm saying is many taker transactions fail to get matching orders and those do not get published. So you cannot backtest only against trade_tape and be faithful to reality

The Dude 2026-07-15 19:43:46

Also do you know about the taker delay?

Also do you know about the taker delay?

trex8791 2026-07-15 19:44:36

Backtesting taker strats is trivial though, just record the book at price_change + book level and simulate latency

Backtesting taker strats is trivial though, just record the book at price_change + book level and simulate latency

Sampoman 2026-07-15 19:44:38

Yeah I know abt taker delay

Yeah I know abt taker delay

Sampoman 2026-07-15 19:45:02

It could be cancels, yes

It could be cancels, yes

Sampoman 2026-07-15 19:46:07

... but still, my gut feeling is that this is sth else

... but still, my gut feeling is that this is sth else

Sampoman 2026-07-15 19:46:24

Though considering how stupidly I acted just 15 mins ago, it could be just me 😄

Though considering how stupidly I acted just 15 mins ago, it could be just me 😄

trex8791 2026-07-15 19:48:15

Do you guys have to consider regime? Or just regular mechanics?

Do you guys have to consider regime? Or just regular mechanics?

Sampoman 2026-07-15 19:48:32

Depends on the algo

Depends on the algo

trex8791 2026-07-15 19:49:46

I find that one hour I can be wildly profitable, yet the next I am not. I'm using black scholes + deribit vol + Binance OBI

I find that one hour I can be wildly profitable, yet the next I am not. I'm using black scholes + deribit vol + Binance OBI

erke 2026-07-15 19:50:15

How do you use deribit vol for 5min markets?

How do you use deribit vol for 5min markets?

trex8791 2026-07-15 19:51:05

I calibrate to the vol smile

I calibrate to the vol smile

Sampoman 2026-07-15 19:52:45

This comes from the fact that the algo itself does not work => it's all probability math playing with you

This comes from the fact that the algo itself does not work => it's all probability math playing with you

Sampoman 2026-07-15 19:53:02

= coin flips piling up

= coin flips piling up

Sampoman 2026-07-15 19:53:42

... and when you find a good algo, it's all about execution

... and when you find a good algo, it's all about execution

Sampoman 2026-07-15 19:53:50

... which is a nightmare on this platform

... which is a nightmare on this platform

trex8791 2026-07-15 19:54:55

Thanks. I'll keep fine-tuning. I'm quite good at execution though. I can share some alpha.

Thanks. I'll keep fine-tuning. I'm quite good at execution though. I can share some alpha.

Sampoman 2026-07-15 19:55:44

I've been playing with this now for like 4 months (?) and tried pretty much everything I can think of. In sim, I have a least 3-4 algos that are profitable against large backtest datasets

I've been playing with this now for like 4 months (?) and tried pretty much everything I can think of. In sim, I have a least 3-4 algos that are profitable against large backtest datasets

Sampoman 2026-07-15 19:55:53

But when I drive them live => they all lose

But when I drive them live => they all lose

Sampoman 2026-07-15 19:56:04

Execute at much worse prices, are too slow etc

Execute at much worse prices, are too slow etc

trex8791 2026-07-15 19:57:58

I have been at this for 7 months. I have 400Gb of backtest data. I understand the frustration.

I have been at this for 7 months. I have 400Gb of backtest data. I understand the frustration.

trex8791 2026-07-15 19:58:34

But I believe if you plot Loss / Volume, you will see constant improvement over time

But I believe if you plot Loss / Volume, you will see constant improvement over time

Sampoman 2026-07-15 19:58:57

My $0.02 is that it's all about execution, nothing else. Anyone can come up with an algo

My $0.02 is that it's all about execution, nothing else. Anyone can come up with an algo