March Madness, Reassessed: The AI Model Performed Well. However, Unpredictable Events Still Occur.

      (NOTE: This article is part of an ongoing series documenting an experiment with using AI to fill the NCAA brackets and observe its performance against years of human expertise. The original article is as follows.)

      A week ago, I shared my experience of entering an NCAA tournament pool using a more structured approach than my usual methods.

      Rather than relying on mascots, vibes, or whichever team looked impressive on Saturday afternoon, I aimed to analyze the bracket like an investor or analyst: distinguishing between raw forecasting and expected value, creating one bracket focused on the highest probability of success, another reflecting pool dynamics, and making choices with some consideration of uncertainty.

      This method resulted in two different brackets. The first was the “most likely” bracket, aimed at maximizing the chances of achieving a good score if the tournament proceeded in a mostly logical manner. The second was an EV bracket for a pool with around 70 participants—not an outrageous contrarian gamble, but a strategy meant to win a real competition instead of simply appearing reasonable.

      So, how did it turn out?

      Overall, quite well, but not without its imperfections.

      The model accurately predicted 13 of the Sweet 16 teams, which is quite impressive in a tournament known for punishing confidence and favoring unpredictability. The fundamental structure of the forecast remained intact. It successfully identified most of the true powerhouse teams and was largely correct about those likely to advance past the first weekend. It generally grasped the overall dynamics of the field.

      However, as March typically does, it also highlighted its weaknesses.

      The most notable misses included Ohio State, Wisconsin, and Florida. Ohio State lost a close game to TCU, 66–64, due to a last-minute layup. Wisconsin fell 83–82 to No. 12 High Point. Florida, the reigning national champion and a No. 1 seed, lost 73–72 to Iowa on a last-second three-pointer. These losses were not slow, obvious downfalls but rather one-possession defeats decided in the closing moments—exactly the type of outcomes that remind us that no tournament model operates in a controlled environment.

      This leads to two possible interpretations.

      One is that the model was flawed.

      The other is that the model was mostly accurate, but single-elimination basketball creates an unreliable scenario for certainty.

      As usual, the truth is a combination of both.

      On the positive side, correctly predicting 13 of the 16 Sweet 16 teams indicates that the underlying framework was valuable. It wasn’t random, merely decorative, or just a more sophisticated way to arrive at the same intuitive guesses as everyone else. It effectively identified quality at a fundamental level.

      On the less reassuring side, the misses also provided insight.

      Reflecting on the process, it seemed to rely too heavily on the notion that “the better team usually advances.” While that may hold true over a season, it is less applicable during 40 minutes in a neutral gym, especially when an underdog can introduce volatility. Wisconsin’s loss exemplifies this point. A more robust upset model might not have chosen High Point to win, but it likely would have viewed Wisconsin as more vulnerable than I did—more prone to a scenario where an underdog heats up from three-point range, stretches the favorite, and turns the final minutes into a gamble.

      Florida’s loss underscores a similar message but at a higher level. A No. 1 seed shouldn’t typically be seen as “likely” to experience an early loss, yet there is a distinction between being robust and being invincible. While the model respected Florida’s capability, it perhaps incorrectly regarded them as safe.

      This distinction is crucial when aiming to win a pool instead of just maintaining credibility.

      Herein lies the interesting aspect of this exercise. In markets, investing, and bracket pools, a significant difference exists between being generally correct and being positioned accurately. A forecast can be thoughtful yet fail to capture where real vulnerabilities lie. The tournament does not offer style points for the best framework if it underestimates the chances of a live underdog getting hot.

      So, what adjustments would I make?

      Not to the core concept. I still believe the ideal way to approach a bracket is to differentiate between highest-probability forecasting and expected-value strategy. Many people unknowingly blend these approaches. They select a champion they believe can win, then make some arbitrary upset choices to “spice things up,” which honestly reflects a lack of a coherent process.

      What I would enhance is the volatility component.

      A more refined approach would place greater emphasis on which favorites are genuinely formidable and which merely appear strong on paper. It would more clearly assess three-point variability, turnover risk, foul trouble, reliance on a single scorer, and how drastically a team’s results can fluctuate from game to game. It would still honor top seeds but remain more skeptical of their invulnerability.

      This distinction has become even more significant now that the original brackets are locked.

      At this juncture, no one can claim they “would have