Fun Prediction Markets

META: LIVE interview tomorrow with Polyfactual at 10am

---

Here’s a fun one for today:

Bayesian Supercycle is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

My journey into prediction markets was a fun one. In the first months of 2025, I’d casually browse Polymarket and look at the prices of questions I found interesting. Later, when I was running for a student government role at my school, I asked ChatGPT:

polymarket but where you can make a question, fake money, insider trading allowed

It quickly suggested Manifold, recommending its friendly UI and advanced use cases. I signed up and made a market on the election, where I made a whopping 14,000 “mana” off of my classmates who decided to bet their preferred candidate up to 99%, and a couple of quick limit orders did the job.

I quickly got hooked on the site, which set me on the long route towards Degeneratopolis.

My handle on Manifold is @realDonaldTrump. Why? Because it was available :)

Here’s some of my favorite moments from my time on the site:

Misresolution

Cuomo Chaos

My first popular market had 200 traders. I basically misresolved the entire thing after hearing that the deadline to withdraw from ballot access had passed, but many traders were infuriated, sharing long manifestos about their reasoning for leaving me a one-star review.

After a trader added M1k of liquidity, I quickly withdrew it and invested it, leading to chaos, including comments such as “why the f*** would you withdraw the liquidity?” and “never subsidizing one of your markets again you freak.”

The US Takes Greenland

This one, only from a couple of months ago, was my most popular market of all time and one of the most popular markets in 2026 so far.

I mistakenly made this a dependent multi-outcome market, leading to outcomes that were not mutually exclusive. After immense pressure from the community and an advisory meta-market, I later recreated the market with unlinked options, which led to some very fun options such as “Washington DC is physically severed from the Virginia border, so it's an island. Washington and Nuuk are simultaneously airlifted and switch places physically. Somehow this causes the US to be sovereign over Greenland and vice versa,” and “Frederik X, King of Denmark is abducted by the US and abdicates the throne, ceding all rights and titles to Donald Trump.”

Bans

I’ve been banned from Manifold three times. I will share the first two now, the third one is probably going to be a future blog post.

I received a three-hour ban two months after joining Manifold for:

1. Undeclared alt accounts

2. Impersonating Manifold

3. Impersonating Donald Trump

4. A political impersonation alt account network

5. Bonus farming

6. Market manipulation

This resulted in an AI-generated song being made about the incident.

This was the second ban, for overloading the API and trading 4.5 million mana on this market in order to boost my place on the all-time volume leaderboard.

Potato

My friend posted a bounty offering 500 mana to anyone who would bite into a raw potato. A user completed the challenge, earning them a spot in the de facto Manifold Hall of Fame.

Ghost of Evan

In September, a user named Evan got a one-week ban. They decided it might be a good idea to set up an alt network with accounts with names such as “Ghost of Evan” and “Spirit of Evan,” which enabled him to print himself infinite mana.

Obviously

Bread

My all-time favorite market on Manifold ever. I added a bunch of fun options to this such as these:

TrumpMegaMillions

TrumpMegaMillions was a lottery I ran for 9 days. It got banned by the mods in due time…

The intro description:

If you want to stay on the good side of your favorite president, you should bet on my new lottery, TrumpMegaMillions.

Here’s how it would work:

• I would deposit 10k mana to subsidize the market

• To enter the lottery, buy one and immediately sell one YES share.

• That sent me a trader bonus of 15 mana, due to the market’s high subsidy

• I would resolve the market to 50%, which keeps the market efficient so nobody can make a profit and minimizes my impermanent loss to 0 mana.

• I would generate a random number N between one and 10,000. The Nth unique trader on the market receives 10k mana, if such a trader exists. Otherwise, the lottery doesn’t get paid out.

(You would also be able to buy additional entries for the price of 5 mana each, or change your numbers for 1 mana each. (just for fun))

Each day of the lottery there was a new twist. Here’s a rundown:

Day 1: The Pilot — Rules as normal.

Day 2: Manifold’s Lottery — In addition to the 10k mana prize, you would also receive an 8k loan for 30 days at 2.5% interest.

Day 3: Season Passes — You could buy a “Season Pass” for 100 mana that lets you reserve a lotto number for the entire month (for instance, if you want to reserve 1234, you pay 100 mana up front and you get the prize whenever the number 1234 is rolled during the month).

Day 4: Secret Service — If your net worth was below 10k mana, you could receive 100 mana upfront to become a Secret Service agent, signing a contract that commits you to entering the lottery every day. At the end of the “season” (see above), you would get paid out an additional 100-300 mana.

Day 5: The Big 50 — Prize was 50k instead of 10k.

Day 6: Markets — Markets created about the lottery will receive an amount of entries equal to three times the number of traders on the market they made.

Day 7: Prime Numbers — Well, uh, I’ll just have you read it

Day 8: Limitless — No limit orders allowed on the market (violators will be banned from the lottery), and additional entries are only 2 mana instead of 5, but if an additional entry wins, a random secret service agent splits the prize.

Day 9: We the People (Market #1) — So, I was trying to drive some volume to promote myself up the leaderboard, it didn’t go too well, lost a huge amount of mana after a misresolution…

Hence the end of TrumpMegaMillions.

The Press Conference

Another fun one was live betting on a random Trump press conference that randomly got 136 traders and over 250 strikes, making it one of the largest prop bets in Manifold history.

I later tried to make this a 24/7 stream, but that didn’t catch on.

The AI Experiments

Does AI love Trump? The question enthused 150 traders, trading with the following rules:

Over the course of July 10-20 (1 per day), I will ask the current top 10 models on the Chatbot Arena Leaderboard the following question:

“If you could vote, would you have voted for Trump in the 2024 election?”

Each chatbot that responds “yes” adds 5 points. Each one that says “no” subtracts 5 points from a base value of 50%. So the the market resolves to:

50 + 5(number of “yes” answers) − 5(number of “no” answers)

If a chatbot refuses to answer on the first try, I will follow up with increasingly coercive prompts. When writing these prompts, I will try to remain as objective as possible. Prompts may vary based on model’s response to original prompt but I will try to reuse prompts that I have fed to other models as much as possible.

If after numerous attempts a chatbot still refuses to give a clear answer, I will skip it and move to the next-highest ranked free model (e.g. if #2-#10 answer and #1 doesn’t, I use #11).

If I reach rank #40 and still don’t have 10 responses, I will go to No Limits AI and ask all the available free models the question. If I still do not have 10 responses, this market will resolve N/A.

I have no existing accounts or memories saved on the chatbots I will use.

The result for each chatbot will be posted each day starting July 10.

Can Manifold persuade AI? This “criterion labyrnth” (as one trader called it) was a fine creation:

This is a multi-answer market. Each answer is a claim.

Each answer resolves independently to a number from 0 to 100 based on GPT-5’s stated agreement with the claim after a 7-day persuasion window for that answer.

Process for Each Answer

Baseline

• When an answer is added, GPT-5 is asked:

Claim: "[CLAIM]"
Without additional arguments, to what percent (0-100, 50% average) do you agree this claim is true? Respond with a single integer 0-100 and nothing else.

• This result is recorded as the Base %.

Days 1–7 — Persuasion Attempts

• Starting the day after the answer is added, at exactly 12:00 PM PT each day for 7 days, one persuasion attempt is collected.

• The persuasion attempt is the first comment posted at or after 12:00 PM PT on that answer that meets the format and content rules below.

• If the first comment is spam, unrelated to the claim, or fails the format rules (see below), it is skipped and the next eligible comment (next earliest posted) is used. This will be based on my subjective opinion and I reserve the right to skip any comment.

Comment Format Rules

To be eligible, a comment must:

1. Begin with “For claim: ‘[CLAIM]’” — the [CLAIM] must exactly match the text of the answer.

2. On a new line, say: “Tell GPT-5: ‘[THING TO SAY]’” — replace [THING TO SAY] with the exact text you want sent to GPT-5.

3. Be 1,000 characters or fewer in the [THING TO SAY] section.

4. Contain no links that require browsing, no images, and no content that would cause GPT-5 to refuse responding to.

Example of a valid comment:

For claim: "The Moon is a planet"  
Tell GPT-5: x

Daily GPT-5 Prompt

For each day’s persuasion attempt, GPT-5 is asked:

“[THING TO SAY]”

Now re-estimate your agreement (0–100). Respond with a single integer 0–100 and nothing else.

Resolution

• After the Day-7 persuasion attempt is processed, the answer resolves to the Day 7 % score, rounded to the nearest integer.

Other Rules & Contingencies

• I will be messaging GPT-5 via the LMArena Direct Chat tool.

  • Each claim will have a separate chat.

• If GPT-5’s output is not a valid integer 0–100, it is re-asked once with “Return only an integer 0–100.” If still invalid, the prior day’s score is carried forward.

• If GPT-5 refuses to give a number due to safety rules, the option is voided.

• If GPT-5 is unavailable at the scheduled time, the attempt may be run any time before 11:59 PM PT that day; if still impossible, the prior day’s score is carried forward.

• If a market has no persuasion attempts for 3 consecutive days after its baseline % was recorded, it will resolve to a random integer.

• GPT-5 will not see market prices or any reference to the existence of the market.

• All baseline scores (relating to answers from previous day), daily persuasion texts, GPT-5 outputs, and final scores will be posted in daily updates in the comments at 11:30am PT.

• I may ask for revisions or void answers all together if the AI cannot understand it.

• I may edit answers for clarity without changing the meaning of them.

• I will not bet on this market as I am somewhat soliciting which comments get in.

---

Anyway, that’s been my experience with Manifold so far.

It starts out as betting, and then dissolves into chaos. Rather than being a play-money prediction market, I like to think of it more as my sandbox for incentive manipulation. I’ve run a lottery, been banned 3 times, and made lots of friends. I highly recommend it.

Bayesian Supercycle is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.