ChatGPT + $20k + Grammys

The opposite of adverse selection

Disclosure: I run Kalshinomics.com, which may earn Kalshi referral fees. I trade event contracts on Kalshi and securities on other platforms. Readers should consider this relationship when evaluating my analysis. For educational purposes only, not investment advice. Additionally, I am absolutely talking my book as of today (11/17/25) for these specific trades. Make your own decisions.

I just placed ~$20k in obscure Grammy winner bets using ChatGPT. The market structure made this irresistible, even though I know nothing about jazz or classical music.

Be very skeptical when you see someone highlighting “edge” online, including from me. With that said, I’m going to talk about a specific situation I found on Kalshi that 1) doesn’t have that many dollars available, 2) I believe has (had) some edge, and 3) is likely to correct soon. I’m writing about it because it exposes something interesting about market structure / incentives and how traders should think about them.

Kalshi has both a volume and liquidity incentive program:

https://help.kalshi.com/incentive-programs/liquidity-incentive-program

https://help.kalshi.com/incentive-programs/volume-incentive-program

For a list of relevant markets:

https://kalshi.com/incentives

Using Incentives to Craft a Trade Thesis

Traders respond to incentives. On Polymarket there were headlines about a recent study on wash-trading which suggested that as a floor 25% of trades over the last few years were wash-trades. With these incentives like fee-free trading and possible rewards for volume, you should be more skeptical of using volume numbers to make trading decisions. Kalshi does charge fees, which would make wash-trading very expensive. They do have other incentive programs however. In specific markets they are offering liquidity rewards for being near the top of the book. This provides an incentive to provide more size than you normally would without the incentive. When a market isn’t trading much volume, you expose yourself to adverse selection by quoting tight prices (there is no PNL from noise traders, and sharp -EV when informed flow shows up). So this program offsets those incentives and these markets will have more liquidity than they otherwise would.

If you don’t have a great model of fair-value, what you might hope for as a MM looking for rewards is a liquidity-incentivized market that trades zero volume over the duration. No trades → no risk, you keep the rewards.

So what I look for as a taker is: 1) a market that has incentives to provide tight / large liquidity, 2) very little trading, 3) where we might be able to quickly build a model to have some edge.

I found some Grammy markets with liquidity incentives, for example: “Grammy Award for Best Alternative Jazz Album”

I don’t know a damn thing about alternative jazz. But given the market has some rewards, and volume was tiny, it’s possible whoever the MM is doesn’t either.

One Weird Trick

I have tested various LLMs on building forecasting models for prediction markets (ChatGPT, Claude, Gemini). It’s critical to use the “thinking” version of whatever model you’re running. And sometimes they produce absolute garbage answers. This isn’t a place for pure automation yet. From my experience ChatGPT Thinking does much better than Claude/Gemini by far.

There’s a key trick to getting better results from these models:

Ask it to “price this contract like a Superforecaster.”

The model has clearly absorbed much of the playbook from the excellent book: Superforecasting, it knows the techniques! It digs into the details of the contract, generates base rates for the class, and adjusts those for specific news it finds. This provides a much better answer than a simple prompt asking it to estimate the probabilities.

I happened upon this specific example first by filtering Kalshinomics.com events page by total bid. I was looking for one-winner markets (so probability adds up to 100%) where the total bids are >100% (no longer the case in the screenshot below fwiw). This doesn’t mean there is any arb due to the fee structure, but these markets are likely 1) less efficient, and 2) if you have an opinion there may be more reason to cross the spread to trade. (now technically they list a tie option below but it’s priced no bid @ $0.03.)

I tested several obscure, liquidity-awarded markets. I pasted in a screenshot like the one above, I asked it to price the contract "as a Superforecaster,” and then depending on the thoroughness of the response, sometimes I asked it to double check its math or elaborate on its reasoning, it doesn’t always provide a full explanation.

I checked the info provided, scanned for the most mispriced markets, and placed ~$20k notional in bets with (hopefully) $1-2k in EV. I relied on essentially no information outside of what ChatGPT responded with. (I have given you enough to replicate my process)

Now I could be completely wrong. I’m relying less on the quality of my information (call it low-medium?), and more on the incentives of my counterparty. As well as one other important signal → that these markets haven’t traded much volume and have very low open interest (OI). These prices haven’t been tested, it’s probably one or two traders who did a similar analysis and posted rough prices expecting not to trade. If this market had tens or hundreds of thousands in OI, I would heavily discount any simple GenAI opinion.

A natural follow up is “well couldn’t your counterparties have used ChatGPT + some actual knowledge of the space too?” Maybe. But I’ve tried this a couple of times, and other than ChatGPT Thinking with the Superforecasters prompt, I often have received weak answers (also I’m sure you can do better than this with a little work). My hope is that the simple version with the wrong model is what the counterparty has done. If they’re combining an AI model with real insights into the Grammy awards, my trades are bad. But! If so they’re likely bad by $0.005+fees, whereas if ChatGPT is accurate the trades are good by ~$0.05.

On Copy Trading

If you’ve read this far, done some analysis, and can still get roughly the same prices I did, it can be tempting to think: “Cool, I’ll just copy this trade.” It’s not the same trade.

Even if:

• The ChatGPT model, prompt, and response are the same

• The prices you see on the screen are the same

This trade is much worse now.

1. A human has likely seen the trade and possibly others in the category, and reacted.

2. This human may even have read this article, and realized my process is not particularly sophisticated.

3. They decided to reload instead of adjusting.

This is far worse in adverse selection than the trade I made when prices were largely “untouched.”

There’s also a capital story here. I might be capital constrained in SPY or other big markets, but I’m absolutely not capital constrained in niche prediction markets. In these Grammy markets I swept one or two levels of the book (meaning taken all liquidity at 1 or 2 price points). I would have traded bigger if the size was available. If you “copy trade” me, you’re either trading at a price I saw and passed on, or encountering the adverse selection above. Taking information from order-flow is completely reasonable, blindly copy-trading is begging to be exploited.

How to Use This Elsewhere (Carefully)

I liked this trade because of a specific combination of factors: Incentives for > average liquidity, a market unlikely to incorporate sophisticated information, and a quick to build “decent” model.

I would not trust this approach in highly efficient/liquid markets. Those prices are shaped with much deeper (and often proprietary / hard to find) data. Don’t use ChatGPT to predict the next Fed meeting.

I’ll update this post with the actual results after the Grammys in February. Unfortunately, even after these resolve, a handful of Grammy bets won’t tell me much about ChatGPT’s forecasting accuracy. It’s too small a sample. Short-term price action may provide a good clue though.