Profits with Speculation: PART DEUX – AI, Sports Simulation, and Analytics

Quick update

So things had taken a big turn. Where we last left off, I was going to turn my speculation efforts into a nice little "journey/story" of my AI AGENTS as a staff.

Then I got super serious about it all.

I'll update with more specifics, but it's already much of what i've shown. I'll explain very quickly and will give more details later, or if wanted.

I built a Sports Forecasting Prediction Model, via GPT. The advancements of 5.0 and 5.1 has made this much more powerful.
I have (LLM CREATED) AI AGENTS "operate" the model. The key is that they (the agents) ALSO have their own set of custom parameters.

So, the process is like this, currently

1. Two AGENTS run the Sports Forecasting Model on my selected NBA, and NFL GAMES.
2. Two AGENTS run the model on selected NHL games for SPECIFIC BETTING MARKETS.

then

3. I take those reports, and feed them to a THIRD AI AGENT, that identifies BETTING MARKET ANAMOLIES / OPPORTUNIES.

My models now are based on more market correlations, instead of 'pure stats'. That's essentially the 1-2-3 process.

All models run WITH an additional MISSION STATEMENT/COMPANY model uploaded - so all models/agents are in line with my goals.
I backtest regularly, as if I'm grading each AGENT on their report(s).

Lastly, I also focus on specific markets that the books "ignore", or just kinda 'make up' an odds # based on how miro-niche the market is.

So, essentially I've created a multi-agent 'intelligence/speculation/analytics' agency.

MONETIZATION

So, the monies DOESN'T just come from actual speculation. What "had happened" was that I really didn't want to do any type of service and be responsible for other people's speculative activities. But two things happened.
1) "it works"
2) I've learned how to generate quite a large amount of data with, like, no tools.

So, I figured WHY NOT. But I still couldn't be arsed. So I extended my "consulting services" (as I mentioned my 'main' IM work) to this. I had created a Telegram group with the idea of doing this giant channel - but, again, I can't be arsed.

So, I basically use it (telegram + telegram premium) to share the data be a 'parlay channel etc; but just with about 5 members (or the people that asked me through reddit! - Oh yea, I created a reddit community, too)

I'll probably expan to that "big channel", since, it's a way to monetize this skill - but, for now I'm the workload light.
GPT is an amazing tool. Think of the ideas that you can do with it, based on what I'm doing now.
Oh yea. I made forecasting models based on/for the MARKETS, themselves, too. Which is what I'm going back to do now!

Thanks for reading if you are! THIS IS YEAR TWO, of what, at first, was just me showing how to make a few extra bucks with stats, your brain, and sports.

A quick addendum to the above.
Occassionally I feed the reports into NOTEBOOKLM.

Hey you mentioned you have an telegram group where you share the sports predictions if possible would love to join that group and try my luck as alot of this just went over my head so yeah would love to make some money with your predictions !

Georgia 09 said:

Hey you mentioned you have an telegram group where you share the sports predictions if possible would love to join that group and try my luck as alot of this just went over my head so yeah would love to make some money with your predictions !

Click to expand...

Yes - and perfect moment to mention the combined use of whop.com + telegram integration + linktree for this portion of the operation.

BHW (Bankroll Horizon Watch) Sports Forecasting

I'm going to be like WWE and trademark 'Sports Forecasting'.

Anyway, like I said, there's several people that somehow found me through reddit etc, so I actually cleaned it up a bit and have it going again, quite recently, actually. As mentioned, I started it, then, when I found "it was working", that's when I get "serious" about the thing.

This instrument really does help.

So, this basically "replaced" my Kindle Publishing that I did for quite some time (5+ years)- both in operation and practice; writing and putting information together. But I still do all the writing / consulting stuff I mentioned in post#1.

With the sports forecasting, I usually post about 2-3 forecasts a day, with some suggested props or market lines, thrown in - which is GOLDEN, because, a year ago, when I start running these telegram pages, I had no idea what any of that stuff meant. The page of the channel link that I gave you begins in either on Sunday 7th Dec with Sunday NFL Football , or Monday 8th Dec w/ NBA + NFL.

Since this is all part of the journey - I actually wanted start (or restart) the newest telegram page on 1st Dec but I'm working on a RACE-TO-MARKETS simulator, (this is where you can bet on a team to get to 10-30-35 etc points first] and somewhere I forgot to save my work when transfering data, and I hate doing things twice - so I put off finishing it, whilst attending to other matters.

Then, in the premium telegram page, I usually AIM to post at least 10 daily + other stats / info etc throughout. I just aim to make sure it's really qualitiy insights, so that others can benefit from the information that I find.

As I mentioned in my quick one-line post I also incorporate notebook lm into some of the analyses etc. If you don't know what it is, it's Google's AI Notebook that turns document files/pdfs in into audio podcasts, infographics, mind maps, slides. It's quite amazing, really. I have it turn my reports into podcasts. Again, it's quite amazing, really. It's an EXCELLENT learning tool. It's also a good addition to a AI multi agent framework as a sports analytics entity.

Tools / Markets mentioned in this post:
Telegram
linktree
Whop
notebooklm
Chatgpt
Kindle Publishing
WWE
multi agent framework
llm modeling
data science
speculation (Fanduel, etc)

NO PART OF THIS POST WAS AI GENERATED.

Good day

Here's an update now that the Monday Night Football game has completed in quite the dramatic way. Remember, now I've pivoted more toward having creating a TEAM of AI AGENTS each with their own parameters, views, biases, specialties etc.

They analyze the data, via the prompts, using, now a combination of GPT, GEMINI, KIMI 2, and sometimes, DEEPSEEK. So, here's how we did with NFL WEEK 14, and making money online through Sportsbooks:
:

AGENT LI RU XI retains a PERFECT record with his converting parlay leg of S. Barkely o60+yds. The finish saw 122 yds.

AGENT'S TOMAR and RAMOT (they are twins. We unfairly treat them as one) -Kamani Vidal Rushing o35.5 rushing yds was another hit. That brings their record to 4-1.

Finally, AGENT ERMAC had Vidal Rushing o30yds rushing for a hit, but had a S. Barkley receiving yds o2.5 which was a miss. He was 3-0. Still, he finds himself in the elusive 4-1 record club.

Again, all props were derived from updates posted here in this very subbreddit, and of course, the sportforecasting telegram.

So, those are the updates applied. It was a g$$d week:

Now including the NEW results:

1. AGENT LI RU XI — 5–0 (100%)​

The undefeated king. Perfect slate. Zero drift. Zero noise.

2. AGENT ERMAC — 4–1 (80%)​

Started 3–0, finished strongly. Only blemish: Barkley receptions.

3. GORGO DU AMLINTON — 4–1 (80%)​

Elite game-environment reading. Only miss: Stroud passing.

4. TENTON THE GREAT — 4–1 (80%)​

Crushed every macro read. Only miss: Egbuka.

5. BELMONT THE FORENSIC — 4–1 (80%)​

Spot-on forensic map. Only miss: Texans sack ladder.

6. NEXUS THE INTERN — 4–1 (80%)​

The Intern continues punching above his rank.

7. TOMAR & RAMOT — 4–1 (80%)​

With the Kamani Vidal hit, the Twins officially enter the 80% club.

8. PAN HENG XI — 3–2 (60%)​

Three excellent reads; two misses that were thin-margin.

Update - end to journey (which indicates a SUCCESSFUL JOURNEY)
So, over the last many months I've been running the Telegram pages, then stopped and now starting again. It's actually for that reason that I had suddenly STOPPED regulary updating this. It served it's purpose in the following way:

When I first started this Journey (last year - PART I JOURNEY, link in first post), I had no idea what props were or who any players on any team on any sport were. I used to be into sports when I had a corresponding video game to simulate seasons at the same time.

Then AI was released to the masses.

So, the short version is that I begain playing with it (ai), programming it, simulating games, creating metrics, creating agents (since 2023), expanding to different league/sports simulations, leading more towards odds correlations and market edges to, holy wow GPT 5.2 is around the corner! I started this with 3.5.

So, since I'm a writer (Kindle publisher for years), naturally, I found a way to write about the analyses an monetize it. As mentioned (in PART I), my reason for starting the journey was for accountability - and, the writing thing.

So now, I have a system that works, and that I repeat a few times here, and can be easily replicated, whether you want to way to make quick money online (Be good at analysis - AI doesn't actually SPOONFEED [like us, here at bhw])

1. Simulation of games (up to 12 at a time).
2. Take all of the simulations and have a "board" of AI AGENTS provide their analysis and picks.

That's it.

3. Then, I share it.

Sometimes there's a step 4 - if I'm actually in the speculative mood, I'll throw all the reports into NOTEBOOKLM, and listen to a CREATED PODCAST based on the data, in addition to THE CREATION OF INFOGRAPHICS, to either help me, or use for telegram/reddit content.

Alright! Remember, AI is your friend (as an entrepreneur) It, literally, can provide you the opportunity to make money online - or, at least create a custom workflow that works to your strenghts.

I love data + sports. Gambling, not so much. Analysis yes. Going to Casinos, no. Analyzing simulations of real world events, yes. So, this works for my odd specific set of super geek interests.

To continue to follow the simulations etc, the link is in the last post I believe. The main telegram is free, my posting on the reddit page I created, is of course, obviously free as well.

I'm just wondering of how others are using GPT.

Edit*
Please excuse my somewhat sporadic use of ( )

Been following this thread for a while and the multi-agent pipeline you described is genuinely impressive — especially the third agent that hunts for market anomalies across the reports. That separation of concerns (forecasting vs. opportunity detection) is exactly where most people stop short.

Interesting timing on this update actually. I've been working on something similar but in a different sport — greyhound racing in the UK and Ireland. Instead of LLM-based simulation, the core is a trained LightGBM model that processes 23+ features per dog (sectional times, trap draw, weight trends, going conditions, rest days, etc.) and outputs calibrated win probabilities. Betfair BSP settlement data is used for backtesting so the odds are real, not simulated.

What you said about focusing on markets the books "ignore" resonates a lot. Greyhounds are massively under-modelled compared to horse racing or football — the books mostly shade prices off each other, so there's genuine edge to be found with proper feature engineering.

We built a free/paid platform around it — trapstats.co.uk — with Kelly Criterion staking and SHAP explanations so users actually understand why the model likes a particular runner.

Would be curious to hear your thoughts on the backtest validation side — how do you avoid overfitting when your dataset is drawn from a relatively small number of markets? That's been our biggest headache.

trapstats said:

Been following this thread for a while and the multi-agent pipeline you described is genuinely impressive — especially the third agent that hunts for market anomalies across the reports. That separation of concerns (forecasting vs. opportunity detection) is exactly where most people stop short.

Interesting timing on this update actually. I've been working on something similar but in a different sport — greyhound racing in the UK and Ireland. Instead of LLM-based simulation, the core is a trained LightGBM model that processes 23+ features per dog (sectional times, trap draw, weight trends, going conditions, rest days, etc.) and outputs calibrated win probabilities. Betfair BSP settlement data is used for backtesting so the odds are real, not simulated.

What you said about focusing on markets the books "ignore" resonates a lot. Greyhounds are massively under-modelled compared to horse racing or football — the books mostly shade prices off each other, so there's genuine edge to be found with proper feature engineering.

We built a free/paid platform around it — trapstats.co.uk — with Kelly Criterion staking and SHAP explanations so users actually understand why the model likes a particular runner.

Would be curious to hear your thoughts on the backtest validation side — how do you avoid overfitting when your dataset is drawn from a relatively small number of markets? That's been our biggest headache.

Click to expand...

Hi. I find it hard to believe that you've been following this thread for a while, considering you recently joined. But, hey, I was a lurker, too.

Regarding the backtesting/validation side, it's actually the smaller concentration and focus of a smaller number of markets that make this most effective. I've seen extended my agent pipeline to player surveillance

.

trapstats said:

What you said about focusing on markets the books "ignore" resonates a lot. Greyhounds are massively under-modelled compared to horse racing or football — the books mostly shade prices off each other, so there's genuine edge to be found with proper feature engineering.

Click to expand...

With regard to horse racing, I focus on MAIDENs. Then I use a feature Engineering prompt based off of WORKOUT DATA. This gives an advantage since most Maidens don't have much race history.

Again, this is another example of a smaller dataset actually being an advantage.

NemoTheOne said:

Hi. I find it hard to believe that you've been following this thread for a while, considering you recently joined. But, hey, I was a lurker, too.

Regarding the backtesting/validation side, it's actually the smaller concentration and focus of a smaller number of markets that make this most effective. I've seen extended my agent pipeline to player surveillanceView attachment 525132
View attachment 525133
.

Click to expand...

Fair point on the lurker thing — found this thread through search, been reading it for a while before registering.

The smaller dataset advantage you describe is exactly what we found with greyhounds too. Fewer races per dog means less noise to overfit to, but it also means you have to be much more deliberate about feature selection. We lean heavily on sectional times and trap draw tendencies rather than win/loss history precisely because the sample sizes are thin.

The maiden horse racing angle with workout data is smart — using pre-race data where race history is sparse is the same logic. Would be curious whether you weight recent workouts differently or treat them equally in your feature engineering.

trapstats said:

Fair point on the lurker thing — found this thread through search, been reading it for a while before registering.

The smaller dataset advantage you describe is exactly what we found with greyhounds too. Fewer races per dog means less noise to overfit to, but it also means you have to be much more deliberate about feature selection. We lean heavily on sectional times and trap draw tendencies rather than win/loss history precisely because the sample sizes are thin.

The maiden horse racing angle with workout data is smart — using pre-race data where race history is sparse is the same logic. Would be curious whether you weight recent workouts differently or treat them equally in your feature engineering.

Click to expand...

That's awesome, thanks for reading for sure! It's all very interesting!
Actually, the way I developed the very first model, using the ai is actually documented here Profits with Speculation – Horse Racing AI Model Journey of Money Management
ss
So, here we are, GPT 5.4 and all. So I haven't used it quite as much, but it is still very accurate and just a bit more updated, in the form of some backtesting. however, it seems to be incredibly accurate. Unlike Baseball, where there's many different metrics etc and such - here, the formulas seem to be pretty much the same.

For different track conditions at time of race AND work out history is factored in and weigh accordingly, relative to other deciding metrics. If doing live, I have tools to account for the steamers and such. However, again, with Maiden, monitoring marketing manipulation really doesn't seem to help.

Unless the purse is REALLY low, I like to limit to 5-7 runners, After that it's either 100% spot on, or 1000% miss.