Apply for AI Grants India

Why this problem needs a careful model

Indian football clubs, academies, agents, and analytics teams often work with incomplete statistics, uneven competition levels, and limited public salary or transfer data. That makes player valuation less straightforward than ranking goals or assists. A winger producing in the I-League, a reserve player in the Indian Super League, and a prospect from a state league may have very different levels of visibility and opposition quality.

CatBoost is useful here because it handles mixed datasets well, including categorical fields such as position, league, club, foot, and nationality. It can estimate a player’s expected market value or predict whether a player is likely to outperform a comparable benchmark. The output should support human scouting—not replace medical checks, video review, references, or contract due diligence.

The strongest projects begin with a clear definition of undervalued: a player whose current observable cost or market estimate is materially below the value your club expects to receive over a defined period.

Define the valuation target first

Avoid training a model on a vague idea of “talent”. Choose a target that can be measured consistently:

• Market-value regression: predict an estimated transfer or replacement value.
• Salary efficiency: estimate expected sporting contribution per rupee of annual wage.
• Future contribution: predict minutes, goals created, defensive actions, or points added over the next season.
• Recruitment classification: label players as high, medium, or low priority based on a club’s budget and tactical needs.
• Ranking: order candidates within a position and competition rather than comparing every player together.

For Indian football, a blended target may be more realistic than a single price. Create a score combining projected performance, age, availability, contract status, adaptation risk, and acquisition cost. Keep the components visible so decision-makers know whether a player appears attractive because of performance, low wages, or simply limited data.

Build a defensible Indian football dataset

Collect player-season or player-match records, and define the observation date. This prevents information that became available after a transfer from leaking into the model. Useful fields include:

• Minutes, starts, age, position, preferred foot, and squad role.
• Goals, assists, shots, progressive passes, carries, key passes, tackles, interceptions, aerial actions, and errors.
• Per-90 statistics, possession-adjusted defensive measures, and team-strength indicators.
• League, club, opponent quality, home or away status, and team playing style.
• Injury absences, suspension, contract length, citizenship, language, and relocation factors where lawfully and ethically collected.
• Reported transfer fees, wages, release clauses, and comparable player transactions.

Use multiple sources and record provenance for every field. Public databases can contain duplicated names, inconsistent positions, missing minutes, and estimates presented as facts. Scouting reports and video tags can add context, but should be labelled as subjective inputs. If you are building an AI product around football data, establish the same audit discipline expected in other Indian open-source AI developer projects: documented schemas, reproducible transformations, and clear licensing.

Engineer features that reflect football reality

Raw totals favour players who play more minutes or belong to dominant teams. Use per-90 rates, minimum-minute thresholds, and shrinkage or reliability features. Examples include:

• Age curve features such as age, age squared, and years to peak.
• Position-specific metrics instead of one universal performance score.
• League-strength and opponent-strength adjustments.
• Availability rate and consecutive-season consistency.
• Contract months remaining and estimated acquisition cost.
• Team possession, pressing intensity, formation, and role stability.
• A “data confidence” field showing whether the player has 300 or 2,500 relevant minutes.

Do not normalise every field automatically. Tree models do not require standardisation, and unnecessary transformations can make explanations harder. Treat missingness as information when appropriate: limited data may indicate low minutes, a lower-visibility competition, or incomplete reporting. Never infer sensitive personal attributes from proxies.

Train CatBoost without leaking future information

Install the library and supporting packages:

pip install catboost pandas scikit-learn shap

A basic regression setup looks like this:

from catboost import CatBoostRegressor, Pool
from sklearn.metrics import mean_absolute_error

cat_cols = ["position", "league", "club", "foot"]
features = [c for c in df.columns if c not in ["player_id", "target_value"]]

train = df[df["season"] <= 2024]
valid = df[df["season"] == 2025]

model = CatBoostRegressor(
    loss_function="MAE",
    iterations=1000,
    depth=6,
    learning_rate=0.04,
    l2_leaf_reg=8,
    random_seed=42,
    verbose=False
)

model.fit(
    Pool(train[features], train["target_value"], cat_features=cat_cols),
    eval_set=Pool(valid[features], valid["target_value"], cat_features=cat_cols),
    early_stopping_rounds=80
)

pred = model.predict(valid[features])
print(mean_absolute_error(valid["target_value"], pred))

A random train-test split is usually misleading: the same player may appear in both sets, and future market conditions can influence past-looking features. Prefer season-based validation, club-held-out tests, or rolling-origin evaluation. Compare the model against simple baselines such as median value by position and age, or a league-adjusted linear model.

Turn predictions into an undervaluation signal

After generating predictions, calculate a transparent gap:

valid = valid.copy()
valid["predicted_value"] = pred
valid["value_gap"] = valid["predicted_value"] - valid["reported_value"]
valid["gap_pct"] = valid["value_gap"] / valid["reported_value"].clip(lower=1)

Prioritise candidates with a positive gap, but add safeguards. Require a minimum number of minutes, cap extreme predictions, and separate players by position and league. A large gap may indicate an error, an injured player, a contract complication, poor data quality, or a market that does not reward the player’s style.

Create a shortlist with columns for predicted value, current cost, confidence, comparable players, key strengths, key risks, and required next action. This is more useful to a sporting director than a single opaque ranking.

Explain, test, and monitor the model

Use CatBoost feature importance and SHAP explanations to identify why a player ranks highly. Check whether the model is over-relying on club, league, nationality, age, or a proxy for reputation. Inspect errors by position, competition, age group, gender competition, and data completeness. A model that performs well overall but poorly for players from less-covered leagues can reinforce the visibility gap it was intended to correct.

Monitor performance after each transfer window. Market prices, tactical fashions, league rules, foreign-player limits, and data coverage change. Retrain only after checking drift and preserving a dated evaluation set. Store model versions and prediction timestamps so scouts can distinguish a current signal from an obsolete one.

Teams can also connect the shortlist to structured feedback workflows. For example, a club may use automated user feedback categorization for Indian SaaS as a conceptual model for tagging scout comments into recurring concerns such as pressing fit, injury risk, or adaptation. Keep human feedback auditable rather than silently converting opinions into labels.

Practical deployment checklist

Before using the model in recruitment, confirm that you have:

• A written definition of value and an explicit decision horizon.
• Licensed, provenance-tracked data with duplicate and identity checks.
• Time-based validation and position-specific baselines.
• Confidence intervals or reliability bands, not just point estimates.
• Explainability reports reviewed by scouts and coaches.
• Privacy, employment, and non-discrimination safeguards.
• A post-signing evaluation plan measuring minutes, contribution, cost, and retention.

The objective is not to discover a magical bargain. It is to make a repeatable process for finding players who merit deeper investigation, especially where public attention and data coverage are uneven. CatBoost can provide the ranking engine; Indian football expertise supplies the context that makes the ranking useful.