Can We Really Predict Which Football Players Will Succeed?

Jonathan Feldman

Glen Ridge High School, Glen Ridge, New Jersey, U.S.A.

Keywords:

Football Analytics, Career Prediction, Player Development, Talent Identiﬁcation, Underachievement,

Predictive Modeling, Sports Data Science.

Abstract:

Predicting whether a football player will achieve their projected career potential is a key challenge for clubs

and scouts. This study analyzes the career outcomes of 8,770 players from the European Soccer Database

(2008–2016), using FIFA video game potential ratings as a proxy for projected potential. To account for the

increasing difﬁculty of skill improvements at higher levels, we apply logarithmic scaling when calculating

achievement ratios. Predictive models were trained on two cohorts: players with complete career data and

those with early-career data (up to age 21). Early-career models achieved moderate predictive performance

(ROC AUC = 0.79), reﬂecting the challenge of identifying long-term success based on limited early observa-

tions. SHAP analysis shows that growth trajectory features, including early improvement and development

patterns, contribute more to success predictions than static physical or technical attributes. We deﬁne success

as fulﬁlling projected potential according to the FIFA rating system – a standardized but subjective bench-

mark. While this does not capture all real-world outcomes, it enables large-scale analysis of developmental

trajectories. These results suggest that tracking player development over time provides better guidance for

talent decisions than relying solely on early physical assessments.

1 INTRODUCTION

Predicting a football player’s long-term career suc-

cess is a key challenge for scouts, coaches, and an-

alysts. Early assessments of potential inﬂuence re-

cruitment decisions, training investments, and trans-

fer valuations. Despite growing access to player data,

many highly rated young players do not reach their

projected potential. In our dataset, only 48.6% of

players with full career trajectories achieved at least

95% of their projected potential, highlighting the dif-

ﬁculty of early talent forecasting.

Explaining why some players fulﬁll expectations

while others underachieve remains an open problem

in sports analytics. Most existing research focuses on

short-term performance metrics or static evaluations,

often ignoring the combination of demographic, phys-

ical, technical, and psychological factors that shape

career outcomes.

This study analyzes these factors using the Euro-

pean Soccer Database (Mathien, 2016), which con-

tains detailed player and match data from major Eu-

ropean leagues between 2008 and 2016. Player poten-

tial scores in this dataset are drawn from FIFA video

game ratings, providing a consistent, though subjec-

tive, proxy for projected potential and enabling longi-

tudinal analysis of player development.

We examine which player attributes are most as-

sociated with fulﬁlling or failing to meet projected

potential. Using correlation analysis and predictive

modeling, we assess how factors such as age, posi-

tion, technical skills, and mental attributes relate to

long-term career outcomes. This analysis contributes

to a better understanding of player development pat-

terns and supports efforts to improve data-driven ap-

proaches in scouting and talent management.

Throughout this study, we use the term football,

also known as soccer in some regions, to refer to the

sport governed by FIFA.

2 RELATED WORK

Talent identiﬁcation and player development have

long been key challenges in sports science and ana-

lytics. Traditional scouting practices typically empha-

size early physical maturity and observable technical

skills, but these approaches often fail to predict long-

term success accurately. (Meylan et al., 2010) and

(Sarmento et al., 2018) highlight the importance of

Feldman, J.

Can We Really Predict Which Football Players Will Succeed?.

DOI: 10.5220/0013671000003988

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2025), pages 38-49

ISBN: 978-989-758-771-9; ISSN: 2184-3201

considering physiological, psychological, and tactical

characteristics in player evaluations, while (G

ullich,

2014) show that early specialization may lead to un-

realistic expectations and suboptimal developmental

outcomes.

Broader reviews by (Vaeyens et al., 2008) and

(Reeves et al., 2018) stress that effective talent iden-

tiﬁcation should incorporate sociological factors and

developmental variability rather than relying solely

on early performance metrics. (Dugdale et al.,

2020) provide longitudinal evidence that success at

the academy level does not guarantee professional

achievement, underscoring the need for models that

account for developmental trajectories.

With the increasing availability of player data, ma-

chine learning methods have become popular in foot-

ball analytics. Schumaker et al. (Schumaker et al.,

2010) explored early applications of predictive mod-

eling for sports outcomes, while more recent studies

by (Decroos et al., 2019) and (Power et al., 2017)

introduced sophisticated frameworks for valuing in-

dividual player actions and decision-making using

match event and tracking data. (Pappalardo et al.,

2019) also contributed a standardized dataset to sup-

port reproducible spatio-temporal analysis in football.

In parallel, research has increasingly emphasized

model interpretability in sports prediction. (Ribeiro

et al., 2016) and (Lundberg and Lee, 2017) estab-

lished foundational approaches to model explainabil-

ity, which have since been applied in football set-

tings to increase transparency in player evaluation.

(Molnar, 2022) offers a comprehensive guide to inter-

pretable machine learning, while (Tavana et al., 2013)

proposed a fuzzy logic framework for holistic player

evaluation.

(Gudmundsson and Horton, 2017) provided a

broad survey of spatio-temporal analysis techniques

in team sports, highlighting the importance of contex-

tual player behavior. Meanwhile, (van Arem et al.,

2025) applied explainable machine learning models

to forecast both player development and market value,

demonstrating how predictive techniques can support

strategic decision-making. Similarly, (Baouan et al.,

2022) investigated which performance indicators are

most predictive of future player value, reinforcing the

relevance of longitudinal data in assessing career po-

tential.

The role of AI in scouting and player develop-

ment is also gaining traction in media and industry

reports (Vicente, 2024; Guardian, 2025), suggesting

a broader shift toward data-driven decision-making in

football organizations.

Despite these advances, few studies have di-

rectly investigated whether players fulﬁll externally

assigned career potential ratings, such as those pro-

vided by the FIFA video game series. These rat-

ings, although subjective, are widely used in foot-

ball analytics due to their consistency across seasons

and player cohorts (Mathien, 2016). Building on this

line of research, our study analyzes over- and under-

achievement patterns using these potential scores and

examines which early-career features most effectively

predict long-term outcomes.

2.1 Datasets

2.1.1 Data Source

We use the publicly available FIFA Player Dataset

compiled by Hugo Mathien (Mathien, 2016), which

contains player data from EA Sports’ FIFA video

game series for the years 2008 to 2016. The dataset

includes player demographics, physical attributes,

and skill evaluations across multiple seasons. The

overall rating and potential scores are based on sub-

jective assessments from the game’s expert panels,

representing perceived current ability and projected

career potential.

The dataset covers approximately 11,000 players

and over 300,000 player-season records, providing a

large sample for analyzing player development pat-

terns. These ratings reﬂect subjective perceptions and

should be treated as estimates of expected career suc-

cess rather than objective performance measures.

We primarily use two tables from this dataset:

• Player: Contains static information such as birth

date, height, and weight.

• Player Attributes: Contains time-stamped

records of player ratings and skill attributes,

including overall rating, projected potential, and

technical and physical skill scores.

While the FIFA video game series is primarily

a commercial entertainment product, its player rat-

ings are informed by extensive scouting networks,

expert panels, and real-world performance observa-

tions. These ratings have become a de facto standard-

ized resource in football analytics due to their broad

coverage, longitudinal consistency, and accessibility

(Baouan et al., 2022; Mathien, 2016). Although sub-

jective, they provide a uniform proxy for projected

player potential across multiple seasons and cohorts

– attributes that are rarely available in open-access

datasets. As such, they are widely used in both aca-

demic studies and industry analyses to explore pat-

terns of player development and market valuation.

Can We Really Predict Which Football Players Will Succeed?

2.1.2 Preprocessing

To ensure sufﬁcient coverage for modeling career

trajectories, we retained players with at least eight

recorded ratings. Player age was calculated dynam-

ically at each observation, and data was limited to

ages 16 to 40 to focus on active professional careers.

The ﬁnal dataset included 8,770 players. Of these,

2,212 had complete longitudinal records suitable for

modeling ﬁnal career outcomes, and 4,683 had at

least three recorded ratings before age 21, supporting

early-career prediction experiments. These groups are

not mutually exclusive; some players appear in both

cohorts.

Final predictive analyses were conducted using

players with complete data for all engineered features

and target variables to ensure consistency and relia-

bility. For binary classiﬁcation, players were labeled

as achievers if their ﬁnal achievement exceeded 95%

of their projected potential; those below this threshold

were classiﬁed as underachievers. This cutoff strikes

a balance between a strict deﬁnition of success and

maintaining a sufﬁcient number of positive cases for

modeling.

While this achievement ratio offers a quantiﬁable

proxy for career fulﬁllment, it remains a subjective

and indirect measure. Speciﬁcally, it does not neces-

sarily reﬂect real-world career accomplishments such

as international caps, top-league appearances, or ma-

jor tournament victories. Therefore, although useful

for statistical modeling, this metric should not be in-

terpreted as a deﬁnitive indicator of career prestige or

professional impact.

2.2 Feature Extraction

We derived over 40 features capturing player devel-

opment and skill progression.

Adjusted Achievement Ratio Over Time. To cap-

ture player development over time, we computed an

adjusted achievement ratio at each observation rather

than relying only on ﬁnal ratings. This metric tracks

how players progress toward their projected potential

throughout their careers.

The adjusted achievement ratio at time t is deﬁned

as:

Adj. Achv. Ratio

log(1 +Rating

− B)

log(1 +Potential

− B)

where:

• Rating

is the player’s overall rating at time t,

• Potential

is the player’s projected potential at

time t,

• B is a baseline value set to 30, representing the

minimum professional-level rating.

We excluded cases where Potential

≤ B to avoid

invalid values in the logarithm. The logarithmic scal-

ing accounts for diminishing returns, where improv-

ing from a rating of 40 to 50 is easier than improving

from 80 to 90.

For each player, we summarized this ratio using

the following aggregate features:

• Mean adjusted achievement ratio across all time

points.

• Maximum adjusted achievement ratio reached

during the career.

• Final adjusted achievement ratio at the last

recorded rating.

• Achievement growth trend, calculated as the slope

of a linear regression over the adjusted ratios.

Additional Feature Groups

• Growth Metrics: Early growth (up to age 21), late

growth (post-21), total growth, growth rate, max-

imum one-year improvement, and growth volatil-

ity.

• Achievement Ratios: Both linear and logarith-

mic versions measuring how closely a player ap-

proached their projected potential.

• Skill Aggregates: Mean, maximum, and ﬁnal

recorded values for stamina, sprint speed, drib-

bling, ﬁnishing, and strength.

• Physical and Categorical Attributes: Height,

weight, preferred foot, and work rate preferences.

Categorical attributes were encoded using label

encoding, and numerical features were standardized

using Z-score normalization to ensure comparability

across features.

Table 1 lists the ﬁnal set of features used in model

training. These were selected from the latest avail-

able player attributes after removing highly correlated

or leaky variables. Categorical features were label-

encoded and all numeric values were standardized us-

ing Z-score normalization.

2.2.1 Feature Correlation Analysis

The correlation heatmap in Figure 1 shows several

patterns in the relationships between features:

• Strong Positive Correlations: The early achieve-

ment ratio and ﬁnal achievement ratio show

strong positive correlations with career progres-

sion metrics such as growth rate, ﬁnal rating, and

strength ﬁnal. This supports their relevance for

modeling career outcomes.

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

Table 1: Final Features Used in Modeling.

Feature Type / Description

Age Numeric (latest observation)

Preferred Foot Categorical (left/right)

Att. Work Rate Categorical (high/med/low)

Def. Work Rate Categorical (high/med/low)

Vision Numeric (passing awareness)

Aggression Numeric (duel intensity)

Positioning Numeric (attack positioning)

Acceleration Numeric (speed buildup)

Sprint Speed Numeric (top speed)

Stamina Numeric (fatigue resistance)

Strength Numeric (physicality)

Dribbling Numeric (ball control)

Finishing Numeric (shot accuracy)

Short Passing Numeric (pass accuracy)

• Athleticism Attributes Show Weak Correlations

with Long-Term Success: Attributes related to

speed and acceleration (e.g., sprint speed, accel-

eration) have weak or even negative correlations

with achievement ratios and career milestones.

This suggests that early athleticism alone may not

strongly predict long-term success, possibly be-

cause these physical traits peak early and do not

directly reﬂect technical or tactical development.

• Technical and Tactical Skills Show Stronger

Associations: Features such as vision mean,

short passing mean, and positioning mean show

moderate to strong positive correlations with

achievement metrics. These skills appear to con-

tribute more consistently to long-term player de-

velopment.

• High Overlap Between Aggregated Features:

Many features were recorded as mean, max, and

ﬁnal values for the same skill (e.g., stamina,

strength, dribbling), leading to strong correlations

often exceeding 0.95. To reduce redundancy and

avoid instability in modeling, we excluded these

overlapping variants and retained a simpliﬁed fea-

ture set. Only non-redundant, numeric features

were used in our classiﬁcation experiments after

dropping potential sources of leakage.

• Growth Patterns and Final Achievement:

Max 1yr growth and growth volatility show

moderate positive correlations with ﬁnal ratings

and achievement ratios. This suggests that

players with late physical development or more

variable growth trajectories can still achieve

high career outcomes, potentially reﬂecting late

specialization or delayed maturity.

• Negative Correlations with Total Growth:

Total growth shows negative correlations with

several performance-related features. This may

indicate that players requiring large improve-

ments to reach their ﬁnal ratings started from

lower initial ratings, complicating interpretations

of their development trajectories.

2.2.2 Descriptive Analysis of Key Features

Table 2 summarizes key features related to player de-

velopment and career outcomes. On average, players

gain 13.9 rating points over their careers, with most

of that improvement occurring before age 21. Final

ratings show substantial variability, reﬂecting diverse

career trajectories.

These patterns suggest that early development

may play a central role in determining long-term out-

comes and motivate further modeling of how phys-

ical, technical, and cognitive attributes inﬂuence ca-

reer progression.

Table 2: Descriptive Statistics of Key Features.

Feature Mean Std Dev Min Max

Early Growth 8.24 5.10 -3.0 20.0

Late Growth 5.67 4.85 -2.0 18.0

Growth Rate 1.20 0.65 0.0 3.5

Stamina Mean 65.8 10.5 40.0 90.0

Strength Max 70.3 15.2 30.0 95.0

Final Rating 72.5 5.40 50.0 93.0

Figure 2 shows the distributions of selected player

attributes used in modeling:

• Physical Attributes: Acceleration (Figure 2a) is

right-skewed, with most players falling between

70 and 80. Extremely high acceleration is rare,

and its marginal beneﬁt may diminish beyond

a certain level. Height (Figure 2b) is approxi-

mately normally distributed around 180–185 cm.

While not strongly correlated with career out-

comes, height inﬂuences positional roles: taller

players often appear in defensive or goalkeeping

positions, while shorter players are more common

in attacking and midﬁeld roles.

• Growth Patterns: Figures 2c and 2d show that

early growth tends to be modest, with only a small

subset of players improving rapidly before age 21.

Late growth is more limited overall, suggesting

that early development has a stronger inﬂuence on

ﬁnal ratings.

• Achievement Ratios: The ﬁnal achievement ra-

tio (Figure 2e) is concentrated near 1.0 for many

players, but a substantial number fall short of their

projected potential, motivating further analysis of

underachievement.

Can We Really Predict Which Football Players Will Succeed?

Figure 1: Correlation heatmap of extracted features. Strong positive correlations appear in red, and negative correlations in

blue.

• Age and Career Coverage: Figure 2f shows

that most player records occur between ages 24

and 26, consistent with typical peak performance

years. This supports the dataset’s suitability for

analyzing full career trajectories.

• Achievement Class Distribution: As shown in

Figure 2g, the dataset includes a modest class im-

balance, with underachievers slightly outnumber-

ing achievers. This was accounted for in modeling

through class balancing strategies.

This descriptive analysis supports the importance

of early career development and suggests that while

physical traits may inﬂuence role-speciﬁc perfor-

mance, technical and cognitive attributes likely con-

tribute more to long-term success. These observations

informed both the selection of input features and the

modeling strategies used in the next sections.

3 MODELING AND ANALYSIS

We used the subset of 2,212 players with complete ca-

reer trajectories to train and evaluate models predict-

ing ﬁnal achievement outcomes. This included both

regression and classiﬁcation tasks using the full set of

engineered features.

3.1 Regression Modeling

The regression models predicting ﬁnal achievement

ratio were trained on the 2,212 players with complete

longitudinal records, ensuring that all growth-related

features and ﬁnal outcomes were fully observed.

We trained an XGBoost regression model to pre-

dict the ﬁnal achievement ratio using the engineered

features. After controlling for potential information

leakage by removing features directly related to ﬁ-

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

(a) Acceleration Distribution (b) Height Distribution (c) Early Growth Distribution

(d) Late Growth Distribution (e) Final Achievement Ratio (f) Age Distribution

(g) Achievement Class Distribution

Figure 2: Distributions of selected features relevant to player development, physical attributes, and achievement outcomes.

nal outcomes (such as ﬁnal rating, potential, and

early achievement ratio), the model achieved an R

of 0.48. This reﬂects the inherent difﬁculty of fore-

casting long-term success based on early-career data.

3.2 Feature Importance and

Interpretation

We applied two approaches to interpret the model’s

predictions:

• XGBoost’s gain-based feature importance, which

estimates each feature’s contribution to reducing

prediction error.

• SHAP (SHapley Additive exPlanations) (Lund-

berg and Lee, 2017), which quantiﬁes the

marginal impact of each feature on model output

across samples.

SHAP Analysis. Figure 3 presents the SHAP re-

sults for the leakage-controlled model.

• Growth-related features such as growth rate,

total growth, and years to peak contributed most

to prediction accuracy. These indicators of devel-

opmental trajectory were more informative than

static skill ratings.

• Physical and technical skills had a moderate in-

ﬂuence. Attributes such as stamina and strength

ranked lower but still contributed meaningfully.

Technical features like vision and positioning had

limited predictive value when trajectory-based

features were included.

Although some technical and physical skills cor-

relate with career outcomes when considered individ-

ually, their predictive value diminishes once growth-

related features are included. This suggests that

player development trajectories capture much of the

variance associated with career outcomes.

The ﬁnal R

result highlights the challenge of

forecasting long-term success from limited early-

career data. Future work should validate these ﬁnd-

ings on external datasets to assess their generalizabil-

ity.

Can We Really Predict Which Football Players Will Succeed?

Figure 3: SHAP analysis results: (Left) Mean absolute SHAP values; (Right) SHAP summary plot showing feature impact

and direction.

Interpretation of Regression vs Classiﬁcation Per-

formance. Although both regression and classiﬁca-

tion models were trained on overlapping feature sets,

their performance metrics are not directly comparable

due to the difference in task framing and evaluation

criteria. The regression model aims to predict a con-

tinuous achievement ratio, which is a more granular

and challenging target, and achieved an R

of 0.48. In

contrast, the classiﬁcation model simpliﬁes the prob-

lem to a binary outcome (achiever vs. underachiever),

which is easier to separate, especially when using en-

gineered features like growth trends. Furthermore, the

classiﬁcation model beneﬁts from thresholding near

the extremes (e.g., 0.95 cutoff), which can lead to

higher ROC AUC scores even if underlying predic-

tions are not highly precise. To quantify variability,

we conducted 5-fold cross-validation and observed

a mean ROC AUC of 0.79 (± 0.02), indicating con-

sistent discriminative performance across folds. This

partial decoupling between the continuous and binary

framing explains why the classiﬁer shows higher dis-

criminative power (ROC AUC ≈ 0.79) despite relying

on similar inputs. This also reinforces the importance

of choosing modeling objectives that align with prac-

tical decision-making goals in scouting and develop-

ment contexts.

3.3 Predictive Modeling and Evaluation

In addition to regression modeling, we trained a

Random Forest (Breiman, 2001) classiﬁer to predict

whether a player would exceed 95% of their projected

potential.

Model performance was evaluated using 5-fold

cross-validation to ensure robustness and mitigate the

effects of random splits. All results reported reﬂect

average performance across the validation folds.

Early-Career Prediction Results. Using only data

available before age 21 and excluding features

that directly encode career outcomes, the model

achieved moderate predictive performance: Accuracy

of 73.95% ± 1.42%, F1 Score of 0.6250 ± 2.14%,

and ROC AUC of 0.7925 ± 1.62%.

These results highlight the difﬁculty of predict-

ing long-term career success based solely on early-

career data. While the model can differentiate be-

tween likely achievers and underachievers better than

chance, predictive accuracy remains limited, reﬂect-

ing the complexity of player development.

Class-Wise Performance. The model demon-

strates higher precision than recall, indicating greater

conﬁdence in identifying players who meet their

projected potential but frequent difﬁculty detecting

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

underachievers. This suggests that the model tends to

overestimate success based on early-career data.

• Precision: 73.08% ± 4.97%

• Recall: 54.73% ± 1.65%

This imbalance highlights the model’s tendency to

favor positive predictions (achievers), while missing

many players who eventually underachieve. Identify-

ing late bloomers and players who fail to reach their

projected potential remains a key challenge.

Error Analysis. False positives typically involve

players who show early promise but fail to improve,

while false negatives are often late bloomers who de-

velop after a slow start.

SHAP analysis conﬁrms that growth trajec-

tory features like growth rate, total growth, and

years to peak are more predictive of success than

static early-career attributes. However, identifying

late bloomers remains difﬁcult.

Despite moderate ranking ability, the model strug-

gles to predict underachievement accurately, under-

scoring the need for longitudinal development track-

ing and caution in early talent assessments.

3.4 Error Analysis: The Limits of Early

Predictions and the Myth of Early

Promise

Despite achieving moderate quantitative perfor-

mance, our predictive model exhibits systematic er-

rors that reveal important limitations in using early-

career metrics to forecast long-term success. Specif-

ically, the model tends to overestimate players who

demonstrate strong early physical attributes and

high initial ratings, while failing to recognize late

bloomers who develop their potential after a slower

start.

False Positives: The Pitfall of Early Promise. Ta-

ble 3 summarizes players who were predicted to

achieve their projected potential but ultimately under-

achieved.

Table 3: False Positives: Predicted Success but Under-

achieved.

Player Name Final Achv. Early Rating Stamina Growth Rate

Anssi Jaakkola 0.93 49.00 43.5 1.50

Lucas 0.94 64.60 58.6 -1.17

David Barron 0.89 53.25 54.0 3.25

Ismael Aissati 0.95 72.00 76.2 0.20

Michele Rinaldi 0.88 66.00 72.0 1.00

The columns in Table 3 and Table 4 provide the fol-

lowing information:

• Final Achv.: The player’s ﬁnal achievement ra-

tio, calculated as actual career performance rela-

tive to projected potential. Values below 1.00 in-

dicate underachievement.

• Early Rating: The average player rating before

age 21, calculated as a simple mean of all avail-

able ratings during this period. This value reﬂects

the player’s observable abilities before reaching

maturity and is used directly as a predictive fea-

ture without transformation.

• Stamina: Average stamina rating during early ca-

reer (before 21), indicating physical endurance.

• Growth Rate: The average year-over-year

change in player rating before age 21. This is cal-

culated as:

Growth Rate =

∑

i=1



Rating

i+1

− Rating



(1)

where N is the number of consecutive rating ob-

servations before age 21. Negative or near-zero

values indicate stagnation or decline in early-

career performance.

Despite strong early ratings and stamina, these play-

ers underachieved due to limited or negative growth,

demonstrating the model’s tendency to overestimate

future success based on early promise without ac-

counting for sustained development.

False Negatives: Overlooking Late Bloomers.

We deﬁne late bloomers as players who show limited

or negative growth before age 21 but experience sig-

niﬁcant improvement in their early to mid-twenties.

Table 4 shows players who were predicted to under-

achieve but ultimately fulﬁlled or exceeded their pro-

jected potential.

Table 4: False Negatives: Predicted Failure but Succeeded.

Player Name Final Achv. Early Rating Stamina Growth Rate

Konstantin Engel 1.00 48.33 48.67 0.00

Krisztian Nemeth 0.99 69.00 63.40 0.00

Kanu 1.00 64.50 60.00 0.00

Jakub Rzezniczak 1.00 61.67 77.00 0.00

Maciej Korzym 1.00 58.75 62.00 -0.40

While these players fulﬁlled their projected po-

tential according to the dataset metrics, only a sub-

set achieved international recognition. Notably,

Nwankwo Kanu became a globally celebrated player

Can We Really Predict Which Football Players Will Succeed?

with major titles at both club and national levels. Oth-

ers, such as Kriszti

an N

emeth and Jakub Rze

zniczak,

enjoyed successful domestic careers or moderate in-

ternational appearances but did not reach global star-

dom. This distinction makes clear that meeting pro-

jected potential does not always translate to elite-level

success and highlights the complexity of deﬁning and

evaluating ”success” in football careers.

These cases illustrate the challenge of predicting suc-

cess for late bloomers. Despite modest early-career

proﬁles and limited growth, some players achieved re-

markable careers, emphasizing the importance of fac-

tors beyond early observable performance.

Fulﬁlling projected potential (Final Achv. ≈ 1.00)

reﬂects meeting expectations based on statistical pro-

jections, but does not necessarily equate to achieving

stardom or international fame. For example, while

Konstantin Engel fully met his projected potential,

his career remained largely within lower-tier German

leagues. In contrast, Nwankwo Kanu, also classiﬁed

as a false negative, became an internationally cele-

brated player despite modest early metrics.

These cases demonstrate that achieving statistical

success according to projected metrics does not al-

ways align with real-world career prestige. This rein-

forces the importance of incorporating broader qual-

itative factors, such as psychological resilience, ca-

reer opportunities, and non-linear development paths,

when evaluating talent.

Correct Predictions: When the Model Gets It

Right. Beyond its misclassiﬁcations, the model also

successfully identiﬁed players who either fulﬁlled or

failed to meet their projected potential. These exam-

ples demonstrate that early-career metrics can be in-

formative predictors when career trajectories follow

more expected development patterns.

Table 5: True Positives: Correctly Predicted Success.

Player Name Final Achv. Early Rating Growth Rate

Ruben Perez 0.99 65.20 2.40

Marco Perez 0.96 70.00 -3.00

Antoine Rey 1.00 41.00 0.00

Michal Svec 0.96 61.00 -0.25

Antonio Candreva 1.00 60.33 1.33

In Table 5, we highlight players who achieved their

projected potential and were correctly classiﬁed by

the model. Notably, Ruben Perez and Antonio Can-

dreva showed strong early ratings and positive or sta-

ble growth. Despite some irregularities, such as nega-

tive growth in Marco Perez, the model correctly iden-

tiﬁed these players as likely achievers.

Table 6 shows examples of correctly predicted un-

Table 6: True Negatives: Correctly Predicted Under-

achievement.

Player Name Final Achv. Early Rating Growth Rate

Mikhail Sivakov 0.85 56.00 0.00

Rafael Dias 0.94 58.75 0.00

Garry Wood 0.85 55.00 0.00

Jamie Mole 0.83 60.50 0.50

Amaury Bischoff 0.88 66.00 0.00

derachievers. These players exhibited moderate early

ratings and minimal or no growth, aligning with the

model’s underachievement prediction. This supports

the model’s effectiveness in identifying stagnating

players early in their careers.

Implications for Talent Identiﬁcation. These ﬁnd-

ings support prior work in sports science that cau-

tions against an overreliance on early specialization

and static performance metrics (Sarmento et al., 2018;

Vaeyens et al., 2008). While early career assessments

offer valuable insights, they often fail to capture the

complex developmental trajectories of athletes. In

particular, identifying late bloomers remains a signif-

icant challenge for data-driven models.

Future work should explore incorporating addi-

tional factors such as injury history, psychological as-

sessments, and changes in coaching environments to

improve predictive accuracy. Ultimately, while ma-

chine learning models can assist in talent evaluation,

they should complement rather than replace expert

judgment and longitudinal scouting efforts.

3.5 Comparing Predictive Value Across

Career Phases

To better understand how the timing of observed data

affects long-term prediction, we compared three ex-

perimental setups: (1) early-career only (ages ≤ 21),

(2) developmental-phase only (ages 22–26), and (3)

a combined pre-peak window (ages ≤ 26). Each

model used the same feature types and binary target

(achiever vs. underachiever based on ﬁnal achieve-

ment ratio ≥ 0.95).

Performance Summary. As shown in Table 7, the

model trained solely on developmental-phase data

achieved the highest F1 score (0.86), outperforming

both early-career (0.63) and full pre-peak (0.77) mod-

els. However, the early-career model achieved the

highest ROC AUC, indicating slightly better class

separation despite lower overall accuracy. These

results suggest that mid-career development offers

stronger predictive signals for identifying achievers,

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

but early patterns still carry useful discriminative in-

formation.

Interpretation. The full pre-peak model beneﬁts

from access to both early promise and mid-career

trajectory, yielding a balanced compromise between

precision and recall. In contrast, the developmental-

phase-only model excels at identifying likely achiev-

ers, but with more false positives, possibly due to

survivor bias: players still active at 22–26 are more

likely to succeed. The early-career model, while less

accurate overall, may better capture cases of under-

achievement and developmental stagnation.

Table 7: Comparison of Prediction Performance Across Ca-

reer Phases.

Experiment Age Range F1 Score ROC AUC

Early-Career ≤ 21 0.6250 0.7925

Full Pre-Peak ≤ 26 0.7746 0.7905

Developmental Only 22–26 0.8596 0.7374

These results highlight the importance of longi-

tudinal observation: while early ratings offer some

signal, the clearest indicators of long-term success

emerge during the professionalization phase. Com-

bining early and mid-career signals yields strong re-

sults overall, but continued development remains the

most decisive factor.

4 DISCUSSION

This study examined whether football player career

outcomes can be predicted using performance trajec-

tories extracted from longitudinal in-game data. We

found that growth metrics observed between ages 22

and 26 were more predictive of long-term potential

fulﬁllment than early-career signals alone. While

combining early and mid-career data improved classi-

ﬁcation performance, developmental-phase indicators

consistently offered the strongest signal for identify-

ing likely achievers.

These ﬁndings support the idea that early static

assessments, such as initial ratings or physical at-

tributes, are insufﬁcient for forecasting long-term out-

comes. In contrast, developmental patterns such as

growth rate, volatility, and late improvements provide

more robust predictive value. This highlights the im-

portance of tracking player progression rather than re-

lying on early promise. It also implies that scouting

systems emphasizing sustained progress during early

adulthood, rather than early peak performance, may

better identify overlooked potential.

It is important to clarify that in this study, “suc-

cess” is operationalized as achieving at least 95% of a

player’s projected potential rating, as assigned in the

FIFA dataset. While this offers a consistent, quan-

tiﬁable proxy for perceived promise, it does not nec-

essarily reﬂect professional prestige, international ac-

colades, or ﬁnancial achievement. Rather, it captures

alignment with expectations as encoded in a widely

used evaluative framework.

Our analysis also revealed systematic model bi-

ases. Players with strong early physical metrics but

little follow-through were often overestimated, while

late bloomers with slow starts were frequently missed.

This reﬂects the inherent limitations of early-career

prediction, even with engineered growth features.

5 LIMITATIONS AND FUTURE

WORK

A key limitation of this study is its reliance on the

FIFA dataset (2008–2016), where projected potential

ratings are based on expert assessments from a com-

mercial video game. These ratings reﬂect perceived

promise, not conﬁrmed professional success, and may

be inﬂuenced by media exposure, reputation, or other

non-performance factors. As a result, our models

predict alignment with subjective expectations in the

FIFA system, not real-world outcomes such as top-

tier appearances, transfer fees, or international recog-

nition.

This distinction matters for generalizability. A

player may meet their FIFA potential but still fall

short of elite professional standards, or may exceed

expectations that were never reﬂected in their early

ratings. While our results shed light on the pre-

dictability of perceived potential, they do not directly

apply to scouting systems based on different goals or

data.

Another limitation is that our models exclude

many contextual factors that can strongly affect a

player’s career. These include injury history, men-

tal resilience, coaching environment, and socio-

economic background. Without such data, the models

may miss key drivers of unusual or nonlinear develop-

ment paths. In addition, the ﬁndings have not yet been

tested on newer player cohorts, so their relevance over

time remains uncertain.

Despite these limitations, the FIFA dataset re-

mains useful for large-scale longitudinal analysis. It

provides structured, consistent player records across

development stages – something that many real-world

datasets, especially proprietary or region-speciﬁc

ones, still lack.

Can We Really Predict Which Football Players Will Succeed?

Future work should extend this analysis to alterna-

tive datasets that track objective performance metrics,

such as club logs, market valuations, or match-level

statistics. Incorporating richer contextual signals and

validating on more recent player cohorts will help as-

sess whether the same developmental predictors hold

across different settings and time periods.

6 CONCLUSION

Our ﬁndings suggest that models incorporating longi-

tudinal development features can moderately predict

whether players will fulﬁll their projected potential.

Growth trajectories, not early static assessments, are

the strongest predictors of future alignment with ex-

pectations.

While standardized ratings such as those used in

FIFA data provide a scalable basis for analysis, they

offer only a partial view of real-world success. Pre-

dictive modeling should be used to complement, not

replace, expert judgment, especially when evaluating

players who may follow nonlinear or delayed devel-

opment paths.

Ultimately, this work contributes to a growing

body of research suggesting that the key to under-

standing future potential lies not in early ratings, but

in how players improve, adapt, and grow across their

developmental years.

ACKNOWLEDGEMENTS

The author gratefully acknowledges the creators of

the publicly available datasets and open-source li-

braries used in this work.

Code and Data Availability

All code and analysis notebooks used in this study

are available at: https://github.com/jonfeld/

icsports2025.

REFERENCES

Baouan, A., Bismuth, E., Bohbot, A., Coustou, S., Lacome,

M., and Rosenbaum, M. (2022). What should clubs

monitor to predict future value of football players.

arXiv preprint arXiv:2212.11041.

Breiman, L. (2001). Random forests. Machine Learning,

45(1):5–32.

Decroos, T., Bransen, L., Van Haaren, J., and Davis, J.

(2019). Actions speak louder than goals: Valuing

player actions in soccer. In Proceedings of the 25th

ACM SIGKDD International Conference on Knowl-

edge Discovery & Data Mining, pages 1851–1861.

ACM.

Dugdale, J. H., Sanders, R. H., and Hunter, A. M.

(2020). Progression from youth to professional soc-

cer: A longitudinal study of successful and unsuc-

cessful academy graduates. Scandinavian Journal of

Medicine & Science in Sports, 30(7):1181–1191.

Guardian, T. (2025). Football coaches could soon be calling

on ai to scout the next superstar. The Guardian.

Gudmundsson, J. and Horton, M. (2017). Spatio-temporal

analysis of team sports – a survey. Pattern Recogni-

tion, 61:491–504.

ullich, A. (2014). Many roads lead to Rome – develop-

mental paths to Olympic gold in men’s ﬁeld hockey.

European Journal of Sport Science, 14(8):706–714.

Lundberg, S. M. and Lee, S.-I. (2017). A uniﬁed approach

to interpreting model predictions. In Advances in Neu-

ral Information Processing Systems (NeurIPS), vol-

ume 30.

Mathien, H. (2016). European soccer database.

Meylan, C., Cronin, J., Oliver, J., and Hughes, M. (2010).

Talent identiﬁcation in soccer: The role of maturity

status on physical, physiological and technical char-

acteristics. International Journal of Sports Science &

Coaching, 5(4):571–592.

Molnar, C. (2022). Interpretable Machine Learning. Lean-

pub, 2 edition.

Pappalardo, L., Cintia, P., Rossi, A., Massucco, E., Ferrag-

ina, P., Pedreschi, D., and Giannotti, F. (2019). A pub-

lic data set of spatio-temporal match events in soccer

competitions. Scientiﬁc Data, 6(1):236.

Power, P., Ruiz, H., Wei, X., and Lucey, P. (2017). Not

all passes are created equal: objectively measuring

the risk and reward of passes in soccer from tracking

data. In Proceedings of the 23rd ACM SIGKDD In-

ternational Conference on Knowledge Discovery and

Data Mining, pages 1605–1613. ACM.

Reeves, M. J., Roberts, S. J., McRobert, A. P., and Little-

wood, M. A. (2018). Sociological predictors in talent

identiﬁcation of junior-elite football players: A scop-

ing review. Soccer and Society, 19(8):1085–1105.

Ribeiro, M. T., Singh, S., and Guestrin, C. (2016). Why

should i trust you?”: Explaining the predictions of any

classiﬁer. In Proceedings of the 22nd ACM SIGKDD

International Conference on Knowledge Discovery

and Data Mining, pages 1135–1144. ACM.

Sarmento, H., Anguera, M. T., Pereira, A., and Ara

ujo,

D. (2018). Talent identiﬁcation and development in

male football: A systematic review. Sports Medicine,

48(4):907–931.

Schumaker, R. P., Solieman, O. K., and Chen, H. (2010).

Sports Data Mining, volume 26 of Integrated Series

in Information Systems. Springer.

Tavana, M., Azizi, F., Azizi, F., and Behzadian, M. (2013).

A fuzzy inference system with application to player

selection and team formation in multi-player sports.

Sport Management Review, 16(1):97–110.

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

Vaeyens, R., Lenoir, M., Williams, A. M., and Philippaerts,

R. M. (2008). Talent identiﬁcation and development

programmes in sport: current models and future direc-

tions. Sports Medicine, 38(9):703–714.

van Arem, K. W., Goes-Smit, F., and S

ohl, J. (2025). Fore-

casting the future development in quality and value of

professional football players for applications in team

management.

Vicente, H. (2024). How AI and Data are Shaping the Fu-

ture of Scouting. Shorthand Stories.

Can We Really Predict Which Football Players Will Succeed?