How I Built a Sports Prediction Engine That Actually Learns From Its Own Mistakes
Most sports prediction tools are black boxes. You get a pick, no explanation, no accountability.
Oracle Odds AI works differently — and I want to show you under the hood.
The core model
Oracle runs a Poisson/Dixon-Coles statistical engine that models the probability of every scoreline in a match. It pulls live odds from four sources simultaneously — ESPN, Pinnacle, Smarkets, and BetExplorer — then calculates expected value by comparing its own probability estimates against what the market is offering.
If Oracle thinks a team wins 60% of the time and the market is pricing them at 50%, that’s edge. That’s a flag.
The AI layer
Raw stats don’t capture everything. Injuries, weather, team momentum, travel fatigue — Oracle passes all of that context to Gemini Flash, which enriches each prediction with qualitative reasoning before a final confidence score is assigned.
The part I’m most proud of: it corrects itself
Every 30 minutes, Oracle automatically settles completed games, compares predictions to actual results, and updates its own calibration. Team strength priors, bias corrections, league-specific weights — all of it adjusts continuously based on what actually happened.
It’s not just a model. It’s a model that gets humbled by reality and responds accordingly.
What I’m building toward
A public accuracy trail by league and sport, so you never have to take my word for it.
Built this solo. Still shipping.
It looks like you are training your model in the process of work, which is commendable. But do you use historical data?
Some historical data, every prediction made gets settled after the game is done and that also becomes historical data. Then recalibration is done and the recalibration bias is store in the db so yes, it’s training as it goes
Really interesting approach — especially the focus on transparency instead of just throwing out predictions. Most tools skip the ‘why’ entirely, so the idea of showing edge vs market + keeping a public accuracy trail is a strong trust angle.
The self-correcting loop is also the part that stands out most — systems that get ‘humbled by reality’ tend to become way more reliable over time than static models.
Also sharing something I’m building in parallel — You have an idea. $19 puts it in a real competition. Winner gets a Tokyo trip (flights + hotel booked, minimum $500 guaranteed). Round just opened, so best odds right now: tokyolore.com
You’ve built something powerful, but the perception needs to shift.
Right now it reads like a system it should feel like an advantage.
“Gets smarter after every match” is what sells here.
If you turn that into a quick before/after or prediction vs reality visual, it becomes much more engaging.
A sharp demo video can help you communicate that instantly I’d be glad to help
Definitely, I’m down
lets connect over LinkedIn or email. please let me know what is your LinkedIn or email i will reach you
Pretty cool idea! So essentially its finding value for games/matches where the market is pricing them in your favor? Definitely a great concept and with all the new sports betting/predictions markets popping up, there will be a lot of demand. Good luck!
Feels weird seeing someone who understood it. Thanks
This comment was deleted 5 months ago
This comment was deleted 5 months ago