Anyone ever thought about using P123 AI to predict single company future earnings or sales revenue and placing bets on the prediction markets? Seems like a simpler and easier hurdle to clear than predicting individual stock price action. The prediction markets are still too illiquid for the large sophisticated AI/ML whales to swim in.
Why? It doesn't seem simpler to me. I'm curious about your reasoning.
Predicting stock price action is essentially what we do when we choose companies to invest in. We're making a probability-based bet that the companies are underpriced. We also do this when we go short, except that we're betting that the price will fall.
Predicting earnings is a somewhat different ball of wax. I've tried many times to come up with a method that would predict company earnings, and I've never come close. So I'm eager to hear your ideas on the subject.
1 Like
I must admit I’m just making an assumption, perhaps a faulty one, because I’ve never attempted to predict future earnings. I assume there are so many more variables that go into price action like current valuations, expectations, macro themes, overall market sentiment, forward guidance, etc. I would have thought predicting future earnings based on company fundamentals would be a more linear projection. Also so many more market participants attempting to predict stock price action that make a more efficient market, so there I would think there should be more dislocations in predicting company earnings to exploit.
But this is the reason why I’m throwing it out there to get feedback rather than jumping into an attempt myself. I’m quite sure there are things I’m overlooking or naive about.
Thinking about these prediction markets more generally…
Early stock markets had thin liquidity, unsophisticated participants, limited information infrastructure, and inconsistent pricing. I think prediction markets today share many of those features.
There is retail-heavy participation, fragmented liquidity across platforms, wide bid-ask spreads, and pricing that often reflects sentiment and availability bias more than actual probability. So, an inefficient market.
I’m not sure what the play is, if any, but it’s very much on my mind.
1 Like
Right now the easy money is prediction markets arbitrage. I am working on being able to start doing this. It might not last long as arbitrageurs join in though. You just bet YES and NO at the same time when it does not add to 1 keeping in mind the fees
Bets are not independent either causing, some inefficiencies in these types of markets.. Easier stated with football as an example. You can bet it will be a low scoring game and that a team that is good on defense will win at the same time (parlays restricted to more independent bets now). Many examples in these bet markets are not independent bets. That is a type of statistical arbitrage FanDuel and others have had to shore up but some are still out there.
You can create a distance from sales target or distance from eps target and try your luck 
Not something I do but that would be a way to start
1 Like
Somewhat related, I've been trying to get my head around what the explosion of AI-driven trading means for running ranking systems in the small and micro cap space. Not doom and gloom, just genuinely trying to think through what changes, what risks emerge, and what opportunities might come with it.
Do the structural protections of the niche hold up? Does anything fundamentally shift? Curious how others are thinking about this.
1 Like
I think a lot of the easy opportunities will eventually vanish but will take a while. Got to make those gains while you can hahah you never know. First of all life is short secondly AI is here. Original ideas will last longer but who knows how long we have once superintelligence is here a few years from now probably before 2035. I might delete some old posts to not help AI. Clock is ticking for easy to spot strategies in my opinion.
As far as the niche (assuming you mean small cap) is concerned it will be impacted since a bot does not care to spend time on micro stocks when it can look at 1000 at a time faster than a person can look at 10. It would just allocate less but still allocate
2 Likes
That is actually a quite interesting idea. Just because you can estimate a earning results correct does not mean you have an edge in predicting the future stock price, we see that over and over again when companies beat expectations. Betting on earnings gives a new dimension of "investing".
There are many many methods to use ML for this. I would start with Gradient Boosted Trees, just because that is what I'm/we are most used to. The key shift is that instead of ranking stocks, we're doing binary classification (beat/miss) or regression on the earnings surprise magnitude.
2 Likes
True, some markets are simple beat/no beat. You could simply specialize a model on yes/no
StarMine makes a business of predicting surprises with ML models: StarMine SmartEstimates. Including the use of unsupervised learning with “RevisionCluster” they say on their web page.
”When the SmartEstimates diverges from consensus (Predicted Surprise) by 2% or more, SmartEstimates are directionally correct 70% of the time”
2 Likes
Factset
Does P123 has access to this granular data ? If yes, with some effort, P123 could create similar product to LSEG.
In theory I think AI should make markets more efficient as it erodes away information advantages, but whether that happens in practice is another question. I believe Cliff Asness/AQR argues the internet counterintuitively has made equity markets more inefficient. The Less-Efficient Market Hypothesis
1 Like
TL;DR: If you predict “sunny with no rain” in the Mohave desert each day you will have a high accuracy score, but that does not make you a good weatherman.
For those who understand that Accuracy can be misleading, 70%, due to class imbalance accuracy for StarMine, may not be so good:
”….they do not publicly publish the Brier Skill Score, Matthew's Correlation Coefficient (MCC), or exact Positive Predictive Values (PPV) in their marketing materials…
…According to a recent LSEG earnings scorecard, in a typical quarter, 71.5% of companies beat estimates, 3% match, and only 25.4% miss estimates. “
Because of this massive class imbalance, if you built a model that literally just output "Positive Surprise" for every single stock, your baseline accuracy would naturally be roughly 70%”
That is not good at all!!!
1 Like
Haha yep its all about how you frame it for marketing.
1 Like
Kind of matches how people are betting
1 Like
If one was going to attempt this might make sense to go for not beats as even if you are wrong most of the time you could still be profitable if you are right 1/3 times. Maybe try that too
This thread got me wondering how good Polymarket itself is at predicting earnings beats. In my opinion a Brier Skill Score (BSS) is the best metric for probabilistic predictions. “There is a 70% chance of rain” is an example of a probabilistic prediction.
Turns out Polymarket has an API and I had OpenAI’s Codex calculate a Brier Skill Score for Polymarket. Results:
”Method summary you can post:
-
Pulled all closed, resolved Polymarket events tagged earnings and filtered to stock questions of the form:
Will () beat quarterly earnings? -
For each market, set:
-
Outcome y from final resolution (Yes = 1, No = 0) using resolved outcome prices.
-
Forecast probability p as the last YES price before market close from CLOB prices-history (14-day lookback window, 1-min fidelity).
-
-
Computed:
-
Brier Score: mean of (p−y)2(p−y)2
-
Climatology baseline using a constant forecast equal to the sample YES rate.
-
Brier Skill Score: BSS=1−BSBSclimBSS=1−BSclimBS
-
Numbers:
-
N = 599 resolved earnings-beat markets (Sep 29, 2025 to Feb 26, 2026)
-
Climatology (YES base rate) used: 0.7295 (72.95%)
-
BS (market forecasts): 0.0981
-
BS_climatology: 0.1973
-
BSS: 0.5026
Optional recent slice:
Since Jan 1, 2026: N=219, climatology=0.6712, BS=0.0801, BS_clim=0.2207, BSS=0.6371”
End of Codex quote.
Conclusion: A perfect BSS is 1.0 A score below zero is possible. Higher is better. For comparison, most weather models often struggle to maintain a BSS less than 0.4. A Brier Skill Score of 0.50 is good in comparison to weather forecasting. Of course, you are not just planning a picnic so you be the judge.
1 Like
Interesting. Did you apply any minimums to polymarket market size? Maybe bigger markets are more reliable than smaller ones. Thanks for sharing
1 Like