Moneyball’s Curse: The World that We Optimize Away

The Oracle of Data

Golazo! Ferran Torres has brought Spain the second star to their badge. The 2026 World Cup delivered no shortage of unforgettable moments, including the end of Joachim Klement’s legendary World Cup winner prediction streak. For three consecutive tournaments, the econometric model of German economist Joachim Klement has correctly predicted the winner of the World Cup: Germany in 2014, France in 2018, and Argentina in 2022.

Ironically, Klement designed this model not to become the oracle of perhaps the largest sports tournament in the world, but rather to, in his own words, show the hubris of economists who think they can forecast anything with a model. Though Spain’s 2026 victory instead of his backed Netherlands ultimately broke his streak, the narrative persists. Algorithms outperform intuition.

Today, this belief extends far beyond football. Data is no longer merely a tool for understanding reality, it has become a symbol of superior decision-making. The prevailing ethos is that if something can be measured, it can–and should–be optimized.

This is the logic behind the modern world’s reliance on data. But have we actually harnessed the power of data properly? Perhaps the answer lies in an unlikely baseball team that practiced a concept called moneyball.

Moneyball’s True Story

So, what is moneyball actually all about? Maybe you’d be more familiar with the concept if you’ve watched the 2011 movie of the same title. Moneyball–the movie–tells the unlikely story of the Oakland Athletics, a baseball team playing in the US Major League Baseball (MLB). Facing budget constraints due to having the third-lowest wage payroll in that MLB season, the Athletics really needed to get creative.

Under General Manager Billy Beane, the Athletics evaluated players to sign based on statistical data rather than traditional scouting intuition. The aim is to exploit market inefficiencies for overlooked traits, most notably on-base percentage (OBP), while bypassing subjective ones such as athleticism and physical aesthetics. The results? Absolutely stellar as the Athletics topped their regional MLB division.

All of this is recorded in the book that becomes the namesake for the method, Moneyball: The Art of Winning an Unfair Game by Michael Lewis. Notice anything? The title of the book itself underlines the ‘unfair game’ of resources that the Athletics were facing. Moneyball is a classic economic case of overcoming scarcity. Unfortunately, the market for high quality players also demands high prices. This is partly because, as Simon (1957) argued, humans have an inherently bounded rationality. As human decision-making faces rationale and information limits, traditional scouts resort to simpler methods of judging a player. Therefore, we overvalue players that meet more easily observable traits while undervalue those who don’t.

Moneyball presents a solution to win this unfair game. At its core, moneyball does not entirely replace human intuition with data, but rather it expands our bounded rationality. Lewis (2003) himself wrote that, “The inability to envision a certain kind of person doing a certain kind of thing because you've never seen someone who looks like him do it before is not just a vice. It's a luxury… When you rule out an entire class of people from doing a job simply by their appearance, you are less likely to find the best person for the job."

The conclusion is quite clear. Data, in the moneyball sense, is meant to uplift humanity. Players and, by extent, employees who weren’t ‘valuable’ in the conventional sense could be given a chance to truly shine in the right environment. A wonderful story with a great message, right?

When Everything Becomes a Number

Moneyball works as a good proxy for today’s world of data analytics. A good concept that when given good implementation would net good results. We can see it everywhere today.

The roots behind such a phenomenon lie, just like moneyball, in information asymmetry akin to Akerlof’s infamous market for lemons. Just like how buyers could not distinguish a high quality car, represented by a peach, and a low quality car, represented by a lemon, in the used car market, employers could not distinguish productive and unproductive job applicants. Therefore, employers rely on what is called job market signaling. First conceptualized by Spence (1973), employers use costly observable metrics–such as prestigious university degrees–as proxies to see through the intrinsic qualities of applicants. The logic is that because meeting these metrics are harder and costlier for lower-productivity applicants, only higher-productivity applicants could realistically meet them.

In this era, using easily observable metrics and proxies is a necessity. Employers can’t be expected to have the time, money, and ability to see the intrinsic qualities of thousands of applicants. This method of judgement is also supported by a quantification of education. The increasing prominence of GPA, standardized tests, and university ranking have provided increasingly conventional signals for employers in filtering job applications. Complementary to this, employers have also started using automated applicant tracking systems (ATS) and AI in hiring to go through the hundreds upon thousands of data coming through.

While the usage of data analytics through observable metrics have much of its merits, the tension also starts to emerge. A number can provide information about a person without telling us everything about that person. GPA may tell us something about academic performance, just as a degree may tell us something about education. But neither can fully capture curiosity, creativity, or unconventional ways of thinking. Going back to the principles of moneyball, we risk losing our ability to envision someone doing something well simply because their qualities do not conform to the numbers we expect to see. Data here is no longer being used in the moneyball sense. Instead of uplifting the overlooked, we risk creating new ways of overlooking them. The question then is not whether these numbers are useful, but what happens when they become the gateway to opportunity.

Goodhart’s Curse

When everything becomes a number, it is important to understand the distinction between a measure and a target. The metrics that we’ve mentioned before are, well… measures. However, we treat them as if they were targets. In doing so, we have broken what is known as Goodhart’s Law. The law states that when a measure becomes a target, it ceases to be a good measure.

GPA is not intelligence, a degree is not productivity, and a university ranking is not educational quality. Yet we use them because they are observable. That in itself is not inherently irrational. According to the World Economic Forum (2025), more than 90% of employers already use some form of ATS and around 88% have integrated AI in that early filtration process. The problem with these methods is that they use a specific cut-off point, such as a 3.0 GPA, and that they are based on past data. In doing so, statistical discrimination is being done.

Statistical discrimination isn’t just a firm level issue, it can transform the labor demographics of a whole nation. In the US, an estimated 27 million people are classified as hidden workers (Fuller et al., 2021). These people are actively seeking employment but systematically filtered out by automated hiring processes. The same study also indicates that 88% of employers acknowledged that qualified candidates are disqualified because they do not match the exact job criteria of the software, not because they are unable to do the job well. Additionally, 75% of resumes never reach a human reviewer and 68% of qualified candidates are filtered out because the system could not read the resume correctly and scored it as incomplete.

These numbers don’t just tell us the problems of our hiring process, but also on what job applicants should do to maximize their chances of getting in. The mechanics is elegantly mapped out by economists Holmström and Milgrom (1991), pointing out a fundamental flaw in how we design incentives. The fact is that real world tasks are rarely one-dimensional. A job usually requires balancing multiple responsibilities, both easily and less easily measured. Their equation proved that when you attach high-stakes outcomes, such as the risk of being filtered out of a job application, to easily measured metrics, human behavior responds rationally, if destructively. People optimize for the ‘numbers’ while leaving much less effort for less quantifiable traits.

A research by Goergen et al (2025), found that job candidates consistently emphasized analytical traits when they believed AI was evaluating them, while downplaying the very human qualities of empathy, creativity, and intuition. This goes beyond the latter stages of job application. From an obsession towards GPA to a focus in university ranking, this phenomenon happens way before we even submit our job application. In a world where your numbers are paramount, the curiosity of knowledge and creativity of thought becomes second to the pursuit of the numbers that are supposed to represent them. It’s either you conform to the conventionally good data or you’re out.

Don’t Become the Statistic

Looking back at the story of moneyball, we find one last revelation. The players were never asked to maximize their OBP. They were just playing baseball and the numbers came afterward. As per Goodhart’s Law, OBP became such a good metric in finding overlooked players exactly because it remains as a measure, not as a target.

Data and metrics are never an inherent problem. In fact, they’ve helped us quite a lot in finding qualified candidates in a more efficient way. The problem is when these metrics unfairly block out and overlook qualified candidates. In that aspect, we are responsible for creating the system and criteria in place. If people begin optimizing the metric, maybe the first question should not be why they are gaming the system, but why did we design a system worth gaming.

Perhaps that is the final lesson of Joachim Klement's model as well. Its greatest achievement was never that it predicted three World Cup winners in a row. It was that a model designed to demonstrate the limits of prediction became a symbol of its power. We should be careful not to make the same mistake with data. The purpose of measurement is not to make the world fit the model. It is to help us see the world beyond our bounded rationality.

References

Amitabh, U., & Ansari, A. (2025, March 28). Hiring with AI doesn't have to be so inhumane. World Economic Forum.

Fuller, J. B., Raman, M., Sage-Gavin, E., & Hines, K. (2021, September). Hidden workers: Untapped talent. Harvard Business School Project on Managing the Future of Work & Accenture.

Goergen, J., de Bellis, E., & Klesse, A.-K. (2025). rAI assessment changes human behavior. Proceedings of the National Academy of Sciences of the United States of America, 122(25), e2425439122. https://doi.org/10.1073/pnas.2425439122

Holmström, B., & Milgrom, P. (1991). Multitask principal-agent analyses: Incentive contracts, asset ownership, and job design. Journal of Law, Economics, & Organization, 7(Special Issue), 24–52.

Lewis, M. (2003). Moneyball: The art of winning an unfair game. W. W. Norton & Company.

Simon, H. A. (1957). Models of man: Social and rational. Wiley.

Spence, M. (1973). Job market signaling. The Quarterly Journal of Economics, 87(3), 355–374. https://doi.org/10.2307/1882010