Trading project
Kalble
Kalble is a market-making bot I built for prediction markets. It reads the live order books on Polymarket, Polymarket US and Kalshi, quotes both sides, simulates its fills against real trades and estimates what each exchange's liquidity reward program would pay it. Everything runs as a dry run, so it never places real orders. I wanted to find out if a small account could actually make money from these reward programs.
How It Works
A market maker posts a bid and an ask and tries to earn the gap between them. On prediction markets that gap is usually small, so the exchanges pay extra to people who keep orders resting near the middle of the book. Polymarket, Polymarket US and Kalshi all have a program like this, and each one scores your orders a different way. The bot's job is to find the market where its share of the reward pool is biggest, quote there and not lose more on trading than it earns in rewards.
I started from Jarrod Watts' open-source jev-trader, which made markets on a crypto order book on Monad. That book had zero bids for minutes at a time when I tried it, so I moved the bot to prediction markets and built the rest on top. That meant adding three new exchanges, a market picker, a simulator for fills and rewards, a loss limit across all the bots, a results page and a small Mac app that starts and stops everything.
The results page splits every dollar into where it came from: rewards, the spread on round trips, inventory moves, fees and luck after the market was already decided. It also shows a 95% range, because a few good sessions can look like an edge when they are really just noise.
What I Found
The trading itself had no edge
The first version made markets on live sports and esports games. Almost all the profit came from reward programs (one League of Legends Worlds game paid about $80 per bot in 4 hours), not from predicting price moves. I ran an AI model against a simple momentum rule and their hit rates were 55% and 53%, with 95% ranges that overlapped, so I could not tell them apart. One game with no reward pool lost $21 per bot.
My own simulator was wrong at first
When I reviewed the code I found that the Polymarket trades feed I was using to simulate fills was cached for about 5 minutes. Fills came in batches at old prices, which made the early profit numbers look much better than they really were. I switched to the live trade websocket and only count a fill if the trade happened after my order was posted. Every session before that fix is marked as stale and left out of the results.
Kalshi's program pays almost nothing at this size
The simulation said Kalshi's Liquidity Incentive Program would pay something, so I placed one real order to check (a $20 NO order on a Senate vote market). It filled, it paid $0 in rewards and the position was down about $9.50 when I checked. Going back through Kalshi's terms showed why. There is a $1 minimum payout per market, both sides of the book have to reach the target size or nobody gets paid that second, and my bot was rotating across 36 markets in 6-hour sessions. Its simulated rewards were spread so thin that no single market reached $1. After I added the minimum to the results, 41 simulated Kalshi sessions earned $0 in rewards and lost $1.65.
Quiet Polymarket markets roughly break even
Long-dated Polymarket markets (things like elections and economic data that settle months out) move slowly, so they are less risky to quote. Over 34 simulated sessions the bot earned $14.24 in rewards and lost $13.61 on the spread, for a net of $0.34. That is basically zero, and I can't trade on international Polymarket from the US anyway.
Polymarket US looked huge until the max spread rule kicked in
Polymarket US has its own reward program with pools that are much bigger than Kalshi's, around $150 to $400 per day for some markets. I added it as a new exchange and wrote the scoring from their docs, with checks against the examples they publish:
score = discountFactorticks from best price x size
walk the book from the best price until targetSize is reached
if maxSpread is set, both sides must reach target within 2 x maxSpread
each side gets half the pool, $1 minimum per market per day
The picker looked at one snapshot of an XRP market and estimated about $41 per day. Over 14 simulated hours the bot actually earned $0.07, or about $0.11 per day. The reason is the max spread rule. That market has a 6 cent max spread, so it only pays when both sides of the book are filled to 500 contracts within 12 cents of each other. Most of the time there was a 5,000-contract bid at 15 cents and almost nothing on the ask side until 35 cents, so nobody got paid in those seconds, including me. One good snapshot made a dead market look like the best one.
Results So Far
| Strategy | Sessions | Rewards | Net |
|---|---|---|---|
| Kalshi LIP (simulated, after the $1 minimum) | 41 | $0.00 | -$1.65 |
| Kalshi real test order | 1 | $0.00 | about -$9.50 |
| Polymarket long-dated (simulated) | 34 | $14.24 | +$0.34 |
| Polymarket US (simulated, started Nov 21) | 3 | $0.21 | $0.21, under the $1 minimum |
Numbers are from November 21, 2026. Sports sessions from before the fill fix are left out. All results except the one test order are simulated.
What I Learned
The biggest thing I learned is that the rules decide everything. Each reward program looks generous from the outside, but a payout minimum, a both-sides rule or a max spread rule can take a small account's share down to zero. I only found those by reading the terms closely and by putting real money in once, and the real test cost me about $9.50, which I think was worth it. The second thing is that my estimates were always too optimistic, because a snapshot of the book doesn't show how often the book is actually in a state that pays.
Next I want the picker to judge a market by how often it paid over the last day instead of one snapshot, and to skip markets where one big order controls the book. If Polymarket US still pays under $1 a day after that, the honest answer is that these programs are built for accounts much bigger than mine.