Likely Sharp

The Kalshi Market for Who Will have a #1 Song on Spotify in April is looking interesting. Morgan Wallen and Ella Langley announced a new song to be released on Friday, the same day as Noah Kahan's new album. Noah Kahan's latest singles this year have done 2.9M and 1.8M streams on day one. Morgan Wallen's 2024 single Love Somebody saw 2.4M streams on Day 1, and Ella Langley's album release earlier this month put her at 2.2M. The market is currently pricing Morgan Wallen at nearly 90% of having a top song, and Noah Kahan closer to 10%. It's surprising to me that Noah Kahan is priced so low considering he's been able to put up similar (and even higher numbers) then Wallen and Langley individually, but will be interesting to see if there's a significant boost from the two together.

Building a poker-related game as a fun side project. What are some interesting poker variants that would be interesting to play for a hand? Ex: Flushes now beat Full Houses, Straights can be made with a gap between them (3, 4, 5, 7, 8), 2 cards are now shown on the river.

Cool arb scenario I found on the MLB Worst Record Markets. Many of these types of arbs are happening all the time, but here's one specific example I found. Within .01 seconds, someone bought 8 shares of No for 19 different MLB teams to have the worst record this season. They spent a total of $143.44 to buy 152 shares of No across the different teams. These shares are guaranteed to be worth at least $144 (+$0.56 profit), since only one team can have the worst record (or n teams tie and the shares are worth $1 / n), and there's even the possibility of the team with the worst record not being 1 of the 19, making the shares worth $152.

@mirrash7 Was trying it out on my MacBook Air so that's probably the difference haha. I like the idea of smaller zones to check too, thanks!

@LikelySharp It might just be the hardware🤷‍♂️ what GPU are you on? Other than that, you can try smaller models, or sectioning off zones so less detections need to happen

My latest computer vision project - Tennis🎾 Featuring: -Player, ball, racquet detection -Player poses -Exclusion zones to kill false positives -Interpolation and smoothing for clean ball trails All running at over 30FPS!

For my sports bettors out there, where would you rank tennis in terms of ability to find an edge? I haven't looked into it before, but it seems like it'd be a great sports to be able to model and endless matches to predict.

At 10M entries, I think the absolute best brackets would be around the 1 in 2B range (chalk or very near chalk). These are about $0.50 per entry. Many people are probably putting in their couple of 10-16 seed upsets which likely put them in the range of 1 in 100s of billions. These are likely $0.005 - $0.01 per entry. And then I bet there are tons that are quite literally near the 1 in quintillions you see from just picking randomly, which make them near worthless. I'd bet you could get a pretty accurate estimate by looking at prior year's bracket pools to understand the average value per bracket based on Vegas odds. With just a quick mental estimate I'd guess that their average bracket EV is somewhere around $0.02 - $0.05 for this tournament, which puts the value of the billion dollar contest around the $200K - $500K range (so less than the $2M in other prizes). Would love to see how this was actually calculated and if the above estimates are near the ballpark.

Here's an exercise: how much is Kalshi's billion dollar contract actually worth? Assume that you have to sell this promotion from SIG's POV

@hanzpo If I go completely random I have a feeling I would have a much lower chance of winning the $1M haha. But hey I'm more than happy to sweat out my $0.61 EV free roll this week!

Another interesting data set from my March Madness research. The "optimal" bracket should have around a 1 in 1.65B chance of winning and would have 0 upsets. Going with the 100th best bracket drops these odds to 1 in 2.1 billion, but likely avoids some chance of duplicating the same bracket as someone else. Going with the 10,000th best bracket drops these odds to 3.2 billion.

@Kalshi I've just built my bracket, and have calculated that the odds of it going perfect are about 1 in 1.65 billion. Thanks for the free $0.61 in EV Kalshi gg 2 ez

The $1 Billion Kalshi Perfect Bracket Challenge $1 Billion for a perfect bracket $1 Million guaranteed to the top scoring bracket $1 Million to charity and scholarships See the full rules and submit your bracket: kalshi.com/billion-dollar… No purchase or deposit required. SIG Parametrics, LLC, a member of the Susquehanna International Group of Companies, is financially backing this promotion.

My favorite March Madness tradition is to simulate brackets until I get the same one twice. It took 3,635,377 simulations to get a match, with this one having Michigan as the winner and all four 1 seeds making it to the Final Four (something I have priced at around 5-6%)

@jspilman @xmayeth Haha I’ve always wondered who is taking the other side of my orders where I’m more confident on the math. Now I’m thinking that a lot of this is just coming from people asking AI

@LikelySharp @xmayeth Your comment has turned this into such a classic Claude moment. So confidently wrong! 😂

I asked Claude to find the most obvious edge on Polymarket. It came back with three bets on the End Of The World and a potential profit of +$18,400. I gave it just one prompt: "find where the crowd is lying to itself the most." Then it went silent for 40 minutes. It came back with a NASA link tracking 34,000 asteroids with meter-level precision. And NASA publishes it for free. The market "5kt meteor strike in 2026?" is priced at 38%. NASA says 0.3%. The market "10kt+ strike?" Market: 20%. NASA: 0.05%. The gap is measured using KL divergence. Anything above 0.05 is an edge. We got 4.76. That is 95x above the threshold. The agent went NO on all three. Resolve is in December 2026. +$18,400 if the world does not end because of a meteor strike. And if it does, we probably will not need the money anyway. But the more interesting question is this: If a 95x gap is visible to anyone with an API, why is the market still pricing it at 38%?

Yeah the math I did here came from the assumption of an average of 1 per year with 81% of a year remaining, but I’m sure that average per year changes depending on the look back period to where 39c is a a reasonable price as well. Anyways, glad to hear that 0.3c sounded unreasonable to others as well since all the previous comments seemed to be saying that the no bet here was a great find

@LikelySharp @xmayeth Yep, and given that almost 3 months have passed in the year with the odds of a 5kt hit within a year, 39c is actually a very efficient market

Ended up trying this yesterday. Set up OpenClaw to periodically check for new Earnings Mention Markets. It then pulls previous earnings transcripts and sends a report on historical hit rates. The market seems to be pretty accurate with pricing in historical hit rates for the more common earnings report words, so the real value seems to come from the ability to pull recent news stories that may be relevant for the given company to say the word/phrase. However, pricing this in then requires pulling previous news stories of similar importance in a similar category for the company, to see how often they talked about that event in the immediately following earnings report which is difficult to automate and has a pretty low sample size, but working on improving this piece now. Another benefit of the automation tool is being able to classify speakers as a company representative vs. someone who is not a company representative (not counted for mention markets)

Has anyone tried using OpenClaw to automate some prediction market research? I know the general "Ask ChatGPT to price this market" is pretty unreliable, but I'm thinking: 1. Take my current code / models and share them with OpenClaw, scheduling a job to run each morning. 2. Share a report with the estimated probability from my model output, along with any mispricings based on current orderbooks for those markets. 3. OpenClaw proactively does research to try to understand why markets may be mispriced and can add that context to the report. I'm thinking the main value here would come from #3, and being able to just quickly pull out model outputs along with research / context gathering that you'd usually have to do yourself.

Did some quick exploration of my recent prediction market win rate by share buy price. The orange dotted line is the breakeven rate (not counting fees). Fortunately I've had a win rate higher than the share price for 94 out of 99 possible full-cent buy prices. For my two worst performing buy prices, I've only won 4.5% of predictions that I got at 6 cents, and only won 58% of predictions that I got at 60 cents. My favorite stat is that I haven't had a loss on a buy >= 92%. Hoping to keep that going!