Overview

Existing AI models for sports gambling primarily focus on numerical analysis and data crunching to inform betting decisions. While this approach demonstrates success, it prompts an inquiry into the potential performance of an AI model employing more human-centric methodologies, such as predictions based on player style reports and team-based strategic analysis. To address this question, I utilized Google Gemini's Deep Research function to generate and analyze basketball players and team reports, culminating in a 6 leg parlay. Although the results may not be entirely definitive, they indicate promising potential for this alternative betting methodology.

Tools

Gemini 2.5 (experimental) and Gemini 1.5
- Large Language Models developed by google
  - Has a mode called deep research which was mainly used to generate reports
Deep Research
- A mode on google AI models that allows them to create research papers on any topic
- The sources covered and range from youtube, blogs, sports websites, stats websites, etc
  - The total websites covered for report generation could easily reach over the 400s
Gemini 1.5 pro
- Used to aggregate reports and make bets
- Has a setting that affects the model output called Temperature
Temperature
- A setting on AI models that allow them to be more “creative“.
  - An AI model's "temperature" setting influences its output. A lower temperature typically yields concise, textbook-like answers. In contrast, a higher temperature can provide a more in-depth explanation, though it carries the risk of generating slightly inaccurate or contextually unrelated information.

Methodology

Based on my own personal knowledge, select the three main drivers of a team's offense
Generate deep research reports on how each player's offense affects the one of the other players for each of the 3 players
Generate a defensive style report on the opposing team
For each player ask the gemini 1.5 pro how the player would perform against the opposing team's defensive style using the previously generated reports as context.
After compiling the final report, Gemini 1.5 pro is used to make parlays based off the final reports
- A parlay for two scenarios is generated, one where the main player does good and where the main player does ok
- The two scenario parlays are generated for the model settings temp 1 and temp 2

Results

Disccussion

The trial's findings indicate that models assuming an "ok" from the main player significantly outperformed others. Among the "ok" models, the more conservative "temp 1 ok" model achieved the best results, aligning with the intuition that calculated ( and somewhat conservative ) risk-taking is best when betting. While "temp 2 ok" currently shows less frequent wins, its large payouts suggest potential for future profitability. If its true win percentage is higher (e.g., 20%), it could become nearly as profitable as "temp 1 ok," and possibly even more so if it succeeds on its highest-odd bets. Conversely, the "temp 1 good" and "temp 2 good" models demonstrate no promise. Although "temp 1 good" can be initially thought to merely require more time due to its higher theoretical risk and variance, its single win came from a bet with significantly lower odds (371) than its average (702) or highest (1838) betting odds. The "temp 2 good" model had no wins, though its relatively high maximum and average betting odds suggest it could potentially become profitable with a larger number of bets.

Conclusion

While not the primary tool for sports betting, an AI betting model based on qualitative reports shows promising results from a small-sample trial. Models predicting "ok" games demonstrate potential profitability with sufficient bets. Conversely, models anticipating "good" games generally appear less viable, with the exception of the risky “temp 2 good" model, which shows limited potential. It's crucial to acknowledge that these findings are based on a small sample size of only 10 bets, necessitating further trials to validate these initial promising observations.