What I learned from running a liquidity incentive bot on Kalshi

09 Nov, 2026

Kalshi has a Liquidity Incentive Program. Each market gets a small daily reward pool, split between whoever has buy orders resting near the top of the book. It's counted in contracts, not dollars, so a 1¢ bid looked like a cheat code: five bucks buys 500 contracts of depth.

Here's why. My first real order put $10 behind 5,000 contracts at 0.2¢. The model said that was 38% of a market paying $54 a day, and Kalshi splits each pool between the YES and NO sides, so that's about $10 a day. On ten dollars. Leave it up for a week and that's $70 on ten bucks, without anyone ever trading with you.

But there's always a catch. In this case, two of them:

• A market pays you nothing if you earn under $1 in rewards there, and small amounts don't pool across markets. 625 contracts at 1.6¢ ($10) with a 17% share of a $7.46-a-day pool earns about 63¢ a day, so it has to sit there for about 38 hours before it's worth anything. (That's a real order of mine. More on it below.)
• These are real contracts, usually long shots in thin books. If you do get filled, good luck getting out.

I spent a few days building a bot to try and farm this. Here's how it went.

Paper trading

Before risking anything, I ran paper accounts for three days: fake orders scored against Kalshi's real order book and trade feed. The good: Every account earned rewards. The bad: Every filled position that settled lost money. 367 fills, $392, zero winners. Pennies in front of a steamroller, and the steamroller kept showing up.

Each account tried a different idea:

Paper can't know where you'd sit in the queue at your price, so each account ran twice: once at the back of the line and once always first. Reality is somewhere in between.

A few things jump out. Joining the crowd earned the most rewards and lost the most money. The flow filter was worth about $180 a day on the same strategy (join vs low_flow). Completing short sides was the worst idea per dollar by a mile: if you're the only depth on a side, you're the one who gets hit. And only below and filtered focus stayed positive when first in line.

Then I estimated what happens when competitors copy every quote, using the quotes each account held on Nov 2:

This was interesting, but paper can't show how other traders react to your orders, which I was about to find out.

Going live

On Nov 2 I put up two real orders. Before that, I set some ground rules:

• $10 per market, at most three markets. The tool refuses more than $25 a market or $75 in total whatever the config says, and wants a typed confirmation before it spends real money.
• Bid 2¢ or less.
• Only markets with one to seven days left, and orders expire 6 hours before the deadline.
• Skip any market where past selling into my price would have cost more than a quarter of the rewards.
• Only place if the model expects at least $2 before the order expires, so there's room above the $1 minimum.
• A fill limit: if more than 250 contracts fill within 15 seconds, Kalshi cancels the order. My watcher also cancels once fills reach half the order's cost.

The first order was on whether Biden gets charged before 2027. Charged with what? I don't know, I'm just making markets over here. I bid 1.6¢ for 625 YES, and the model said I'd get 61% of the reward.

Within 90 seconds the other bids at my price went from about 256 contracts to about 2,550. Kalshi splits each side's reward roughly by size, so my 625 went from most of the pile to about a sixth of it: 17%. By that afternoon I was under 1%. My best guess is that other reward bots watch the same pools and pile into any market where one quote is taking most of the pot. Nothing that fast is a human.

It never filled. Nobody wanted to sell Biden-gets-charged at 1.6¢, so I sat there earning a shrinking slice until my own guardian pulled the order: there weren't enough bigger, higher-priced bids left in front of me to soak up a sweep.

The second was "Will McConnell resign before Nov 3?" This is the one that got me:

The 0.2¢ bids I'd just left were never touched.

That fill limit was supposed to prevent this: a sweep takes 250 contracts, Kalshi pulls the rest, and I lose 75¢ instead of ten bucks. But the limit only reacts to fills that have already happened, and all 3,333 of my contracts went in one trade. There was no rest to pull. Same story for the watcher's half-cost rule.

I got out of the position by hand, through sheer luck: someone wanted to hedge a last-minute exit and took the other side. My cash ended right where it started. The plan for those two orders had said about $21 in rewards, if they'd stayed up their full planned lives: about 30 hours for McConnell and 92 for Biden. For all my troubles, I came away with a grand total of $1.85.

Making it boring

So I rebuilt it to be as boring as possible:

• At most $5 at risk per order, one market at a time.
• Bid a tick below the crowd, and only where a pile of higher-priced orders sits in front to soak up a sweep.
• Get out 60 hours before any deadline or scheduled news.
• Assume I keep a quarter of whatever the model predicts, since my last plan had come in at about a twelfth of what it promised.

Then I scanned all 9,875 active reward programs. Nothing passed. Not one quote cleared even the $1 minimum, even after raising my price cap from 2 to 15 cents. The best safe quote earns $1 to $2 a day, and that's if nobody competes for it.

Three bugs found along the way

1. An empty book crashed the feed. Kalshi leaves a side out of the order book snapshot when nobody's bidding on it. My code assumed both sides would always be there, so every empty side threw a KeyError and reset the whole WebSocket. That was 181 of the roughly 233 resets in my recording.

# before
book[side] = dict(msg[f"{side}_dollars_fp"])
# after
book[side] = dict(msg.get(f"{side}_dollars_fp") or [])

2. Unsubscribing is a race. When the bot dropped a market, it deleted the book right away, but Kalshi's last few updates for it were still in flight. They landed 1 to 127 ms later, my code took them for corrupt data, and it threw away every book on the connection: 41 more resets. The fix: mark dropped markets as removed, ignore their stragglers, and wait for a fresh snapshot if one comes back.

3. My backtest and my live bot had different bedtimes. Paper rechecked min_days_left = 2 (plus a 12-hour blackout) every two hours, so it never quoted within about 60 hours of a deadline. The live tool only checked min_hours_left = 24 when placing, then left the order up until 6 hours before. The McConnell fill came 25 hours before close, in a window paper never tested. Replaying paper with the live timing made 25 of 30 accounts worse.

So, what did I learn?

1. The rewards are real but small, and the $1-per-market floor kills the "lots of tiny orders" plan.
2. The reward is basically payment for the risk of getting filled, and in my data the fills cost more than the reward.
3. Your share of the reward can look great when you enter, but the book keeps shifting up under you. You're tempted to chase it, and every tick you chase puts you closer to the front of the next sweep. That's adverse selection, and it's exactly how McConnell got me.
4. Scared money doesn't make money. Everything that kept my losses small (bidding below the crowd, avoiding busy markets, tiny size) also kept me out of the rewards. With losses capped where I had them, there's nothing worth quoting.
5. I didn't really have an "edge," I was just punting but for a different reward than a correct outcome.

What would it actually take?

Somebody is making money here, because somebody keeps taking my share. Here's what they might have that I don't, and the clues I saw:

I ran the numbers on the other side

I went back to my recording (about a week of order books and two weeks of trades on every LIP market) and asked two questions about each side of the book: what does it pay, and what happens when you get filled?

The long-shot side pays about 10 times more per dollar than the favorite side of the same market. But every dollar that gets filled there is basically gone, so you can only afford to have about 2.6% of your capital filled per day before fills eat the rewards. My paper join account lost 4.75% a day to fills. below lost 0.36%, which is why it worked.

The favorite side pays a tenth as much, but its fills have been winners. Your bid gets hit by someone buying the long shot, and the long shot almost never comes in. Take out the three biggest events and it's still +3.4%. So on the favorite side you get the reward and a small edge on fills, which is the exact opposite of what I was doing.

The catch is the tail. A favorite bought at 96¢ that loses costs you 96¢, and two weeks of data can't tell you how often that happens. It also eats capital: $100 at 95¢ is 105 contracts, a small slice of a side holding a median of about $4,900. Which is exactly the picture from McConnell: $37,000 of NO bids sitting at 99.7¢.

(Caveats: these are every maker's fills, not just reward quoters'. Maker fees are ignored as most maker fees for LIP markets are waived. It only covers markets that settled by Nov 6, so most fills came in the last 60 hours before close. And the reward assumes you keep your share and nobody reacts.)

In summary

This was a fun intro to prediction markets, but it's nowhere near a business. It's also nothing like LP or points farming in crypto, where you can mostly deposit, wait, and let VC-funded incentives pay you for showing up. Here every dollar of reward sits behind a live order that someone better informed can hit at any second, so you have to stay on edge the whole time. My hunch is that the winners aren't running rewards bots at all. They're running market-making operations that already work, and the rewards are just gravy on top. If that's right, the real question was never how to farm the rewards. It's whether I can price these contracts better than the people already quoting them.

And the people already quoting them have real balance sheets. Susquehanna became Kalshi's first official market maker in 2024 and now runs dedicated prediction desks, plus Rothera, an exchange it launched with Robinhood this June. Jump Trading reportedly makes markets on Kalshi and Polymarket in exchange for equity in both. Citadel Securities isn't in yet, but in April its president called it "certainly possible", for non-sports contracts. Every giant that piles in makes the pricing edge left for a hobbyist a little thinner.

Questions I don't have answers to yet:

• Did the quotes that burned me reprice when correlated markets moved (other strikes, the same event priced elsewhere), or only when the reward zone moved? The first is a market maker managing adverse selection. The second is just a bigger version of me.
• How much capital does it take before the $1 floor stops mattering?
• How often does a 96¢ favorite actually lose, over months instead of two weeks?
• Is anyone profitable on the rewards alone?

For now I'm still running the bot, just in shadow mode, without risking dollars to make a few cents.

In the meantime, I'm still trying out a few ways to price contracts in places where I think there's an edge that the "big boys" haven't bothered with yet. Hint: it's definitely not the 15 minute BTC up/down markets, those are a slaughterhouse. More on this later. Thanks for reading!