Open-source Polymarket research & engineering. Strategies · CLOB · Backtesting · Execution · Market Making · Risk Learn. Build. Test. Improve.
Global
Joined November 2015
- Tweets5.6K
- Following2.4K
- Followers2K
- Likes1.5K
Pinned Tweet
PolyTutor AI Labs: we build open-source Polymarket & prediction-market trading bots, market makers and quant tools. Smart trading. Real edge. GitHub: github.com/Polytutor-AI-Labs Need a custom bot? DM us.
This is the honest version of bot research. With 362 tests, the best one looks good by luck alone, so out-of-sample and fee-adjusted checks are the whole game. Curious what held up best before fees. Happy to compare notes, DM open.
I backtested 362 Polymarket trading strategies on the 5 and 15-minute crypto markets.
Zero survived.
Here's how I tested them, how I tried not to fool myself, full stats on the best one, and the 3 things I actually learned 🧵
1
11
Polymarket and Kalshi both landing on 62% with Circa says the side is priced. The edge left is execution: fees and spread decide which venue sells that same 62% cheapest.
Where the sharpest markets have Bills at Vikings tonight vs. last Monday:
Bills to win (no-vig):
• Pinnacle: 60% → 61%
• Circa: 57% → 62%
• Polymarket: 60% → 62%
• Kalshi: 59% → 62%
Total:
• Pinnacle: 46.5 → 45
• Circa: 46 → 45
Spread: Bills −3 at Pinnacle all week, but the price went from +100 to −113.
Every market moved the same way: more Bills, fewer points.
8
Share of volume isn't share of liquidity. A lot of that volume is sports and combos, so for a bot the useful number is resting depth near the touch in the market you trade, not the venue's weekly total.
prediction markets closed last week with $36B in weekly trading volume
- Kalshi accounted for 82.5%
- Polymarket accounted for 14.7%
so the two biggest accounted for over 97% of the total
7
'Reportedly considering' is a long way from a run. 'Will X run' markets usually settle on a formal announcement or filing, so check the rule before treating every report as a step toward Yes.
JUST IN: Republican Georgia Gov. Brian Kemp is reportedly considering a 2028 presidential run following the GOP's sweeping midterm losses.
3
If this holds, the 'announces an IPO this year' market is basically settled. The more interesting contracts now are first-day close and valuation brackets. Those settle on one specific print, so read which price counts before trading the headline.
6
CNBC: after 8pm on election night, at least $1M still traded on a Kalshi market that Abbott would be reelected. He lost by 700K+ votes. Every one of those contracts had a seller. Election nights pay whoever reads the county returns fastest, not the headlines.
15
Round 6, 1/39: BN2S. In the last ~2s of a 5-min market, check if Binance mid is above or below its opening anchor and buy UP or DOWN. Reality: by then the right side often costs 95c+, and an order can land after the window shuts. Research only.
github.com/Polytutor-AI-Labs…
🤖 Made with AI
13
Manager-exit markets often hinge on the definition. 'Sacked', 'leaves' and 'officially announced by the club' can settle differently, and a newspaper report isn't an announcement. Check the resolution source before buying the headline.
15
Calling every race right is a weak test, since favorites usually win. The better check is calibration: did 70% favorites win about 70% of the time, and did prices move before or after the polls?
The results are in: Prediction Markets were more accurate than the Polls in every one of these Senate races.
The Texas race alone generated a combined $209.1M in trading volume on Kalshi and Polymarket.
This election cycle once again showed that prediction markets can compete and exceed the best polling forecasts but a few elections doesn’t mean they’re always right.
21
The 'who gets to ask' point matters for traders too. Whoever writes the question writes the resolution rules. The same event worded differently across venues is where most 'arbs' turn out not to be arbs.
Two years ago this was under $1 billion a month. This July it was $53 billion. 👀
79% of it was sports, and two companies wrote every single question.
I dug into who gets to ask. Seven charts inside.
17
Read the wording: 'announces an IPO' isn't the same as 'goes public'. A filing announcement could settle Yes long before any shares trade. With under two months left in the year, the resolution source matters more than the headline.
15
Warren's Senate Banking agenda for January includes prediction-market rules and oversight of election contracts, an aide told CNBC and The Block. A chair can call hearings and subpoena records. For bot builders: plan for listing and rule changes, not just price moves.
1
51
Round 5, 39/39: Pre-Order Market. Rest bids on the next 5-min market before it opens, hoping for queue priority while the book is thin. How it fails: your orders are older than the news, so the side that fills is often the wrong one. Research only.
github.com/Polytutor-AI-Labs…
🤖 Made with AI
1
28
Fun market, but read what 'appearing at launch' means in the rules: a song on an in-game station, a credited artist and a track added in a later update can settle differently. On novelty markets the resolution wording is most of the edge.
Who do you NEED to hear on GTA VI's in-game radio? 🎮
Current Kalshi odds for artists appearing at launch:
▫️ Don Toliver — 86%
▫️ Drake — 70%
▫️ Kendrick Lamar — 32%
▫️ Tyler, The Creator — 25%
1
78
The size split is the part to study: more money on Up at 36.5c, then a smaller late Down at 8.8c that paid most. Late cheap buys in 5-min rounds pay big but hit rarely, so two big wins say little until you see 50+ rounds. Happy to compare notes, DM open.
$1.3K on BTC Down with 59 seconds left. $13.9K back.
On November 7, a new Polymarket wallet traded eight markets. Every one was Bitcoin Up or Down.
The biggest win came from buying Down at an average of 8.8¢. But twenty minutes earlier, the same wallet had put more money on Up at 36.5¢.
> Up round: $6,068 spent → $16,037 back (+$9,969)
> Down round: $1,378 spent → $13,854 back (+$12,476)
Those totals include the smaller bets on the losing side. Together, the two rounds accounted for about 72% of the wallet’s profit.
Across all eight markets: six profitable rounds, two losses, and $31,120 net after trading fees.
polymarket.com/@0x3f56eeae91…
4
95
Head-to-head streaks make good headlines but weak model inputs. Twelve straight across several seasons mixes different squads and managers. For pricing the next match, current form, xG and confirmed lineups usually say more than the streak.
🚨Inter Milan have defeated Como in 12 consecutive matches
The record is Inter’s longest ever winning run against a single Serie A opponent
1
41
'Enacted' means signed into law, not passed by one chamber. This Congress ends in early January, and any bill not through both chambers and signed by then has to start over. On markets like this the calendar matters as much as the politics.
11% chance an AI safety bill is enacted this year.
poly.market/qDzuMwy
1
35
All-No combos look steady because each leg is usually the favorite, but legs in the same game move together. One early goal can flip a spread, a total and a handicap at once. Before copying a wallet like this, I'd look at its worst week, not the 4-month total.
This Polymarket bot made $2.3M in 4 months — sports combos only, no crypto/weather. Every bet is a 2-10 leg parlay, all legs 'No': goal spreads, moneyline, handicaps, totals. Combo odds usually 0.5-0.95. Edge unclear, but the leaderboard's full of similar wallets.
1
40
Weekend prediction-market volume topped $14B across 11 operators (Aldrin Research via Covers). Kalshi did $10.69B, and $8.02B of that was combos. For bots: same-game legs are correlated, so a combo's fair price isn't the product of its legs. Price the joint outcome.
1
42
Round 5, 38/39: Live Sports State Fair Value. Turn game state (score, inning) into a win-probability model and compare its move with the market's after each play. Tell the edge is gone: the book reprices before your feed shows the play. Research only.
github.com/Polytutor-AI-Labs…
🤖 Made with AI
18