Apply for AI Grants India

Why outlier detection matters in Indian football

Transfer datasets from the Indian football ecosystem are often small, uneven and assembled from multiple sources. A fee may be reported in rupees, euros or not disclosed at all; contract lengths may refer to the initial deal or an extension; and player records can be duplicated across leagues, seasons or spellings. These inconsistencies can distort scouting models and financial analysis.

An outlier is not automatically a bad transfer or a suspicious player. It is a record that is unusually different from the rest of the dataset. That difference may reveal a data-entry mistake, a rare but legitimate signing, an undervalued player, an unusually expensive acquisition, or a market segment that should be analysed separately. Isolation Forest helps you prioritise records for review; it does not replace football judgment or due diligence.

The same workflow is useful for clubs, agencies, analysts and Indian sports-tech startups building recruitment products. Teams already developing data products can also draw on practical guidance from Indian open-source AI developer projects when choosing reproducible tools and deployment practices.

How Isolation Forest works

Isolation Forest detects unusual observations by randomly splitting feature values. Common observations require more splits to isolate because they sit in dense regions of the data. Rare observations are separated quickly and receive a higher anomaly score.

The method is a useful first choice for transfer data because it:

• Handles several numerical features at once.
• Works without labelled examples of fraud, errors or exceptional transfers.
• Scales reasonably well as historical seasons and leagues are added.
• Produces a ranking of unusual records rather than only a binary verdict.

It is less reliable when the dataset mixes fundamentally different markets. A goalkeeper and a striker should not necessarily be compared with the same expectations, and an Indian Super League transfer should not be treated as identical to a youth or semi-professional move. Segment the data by competition, season, position or transfer type before modelling where sample size permits.

Build a credible transfer dataset

Start with a data dictionary and record the source and collection date for every field. Potential columns include:

• Player identifier, age at transfer and primary position.
• Origin and destination club, league, season and transfer window.
• Transfer fee, currency, reported fee status and fee type.
• Contract duration, renewal status and loan or permanent status.
• Appearances, minutes, goals, assists and relevant performance metrics before the move.
• Nationality, domestic or international status, and player valuation if available.

Do not convert missing fees to zero. A free transfer, an undisclosed fee and a missing value represent different situations. Keep a separate fee_status field and either model disclosed-fee records independently or use a carefully designed treatment for missingness. Convert currencies using a documented rate and retain the original amount and currency for auditability.

Clean names and club identifiers, remove exact duplicates, and check impossible values such as negative fees, implausible ages or contracts longer than the competition permits. Avoid leakage: if the model is intended to support a pre-transfer decision, do not include information published after the move, such as later performance or revised market value.

Feature engineering matters more than adding dozens of columns. Useful transformations include log_fee, age bands, minutes per appearance, fee relative to club wage or revenue where available, and an indicator for loan, free, undisclosed or permanent transfers. Apply a log transformation to heavily skewed fee data, but preserve the raw value for interpretation.

Implement the model in Python

The following example uses disclosed fees and a small set of numeric features. In a production pipeline, add validation, segmentation and versioned preprocessing.

import pandas as pd
from sklearn.ensemble import IsolationForest
from sklearn.impute import SimpleImputer
from sklearn.pipeline import make_pipeline

transfers = pd.read_csv("indian_football_transfers.csv")

# Keep only records suitable for this fee-focused analysis
model_data = transfers[
    (transfers["fee_status"] == "disclosed") &
    (transfers["transfer_fee_inr"] >= 0)
].copy()

model_data["log_fee"] = (model_data["transfer_fee_inr"] + 1).apply(__import__("math").log)
features = ["age", "log_fee", "contract_years", "prior_minutes"]

pipeline = make_pipeline(
    SimpleImputer(strategy="median"),
    IsolationForest(
        n_estimators=500,
        contamination="auto",
        random_state=42,
        n_jobs=-1
    )
)

pipeline.fit(model_data[features])
model_data["anomaly_score"] = -pipeline[-1].score_samples(
    pipeline[0].transform(model_data[features])
)
model_data["is_outlier"] = pipeline.predict(model_data[features]) == -1

review_queue = model_data.sort_values("anomaly_score", ascending=False)
print(review_queue[["player_name", "anomaly_score", "is_outlier"]].head(20))

For clearer production code, fit the preprocessing and estimator as separate named objects or use a custom pipeline that exposes scores consistently. The important outputs are the anomaly score, the model version and the features used—not merely a Yes/No label.

Tune and validate the results

contamination represents an expected proportion of anomalies, but there is rarely a defensible universal percentage for Indian football. Avoid setting it to 0.01 simply because it is a common example. Begin with contamination="auto", inspect the top-ranked records, and compare results across plausible settings such as 0.02, 0.05 and 0.10.

Use several checks before acting on a flag:

• Compare the record with peers from the same season, position, competition and transfer type.
• Verify the fee, contract and player identity against the original source.
• Run multiple random seeds and examine whether the record remains highly ranked.
• Train on earlier seasons and test whether the same patterns appear in a later window.
• Compare Isolation Forest with robust z-scores or a local-density method.
• Ask a scout, sporting director or finance lead to review the shortlist.

A useful evaluation measure is review precision: of the top 20 records sent to analysts, how many were genuine data issues or commercially important cases? This is more actionable than claiming that an unsupervised model has perfect accuracy. Keep a review log so human decisions become labelled data for future supervised models.

Turn anomalies into football decisions

A high score should create a review task, not an automatic rejection. Classify each flagged transfer into categories such as data error, exceptional fee, unusual age-value combination, contract anomaly, potentially undervalued player or legitimate strategic signing.

For scouting, combine the anomaly score with performance, availability, injury history, tactical fit and expected salary. An unusually low fee may indicate opportunity, but it may also reflect injury, contract expiry or incomplete reporting. For finance teams, compare the fee with amortisation, wages, agent costs and likely resale value. For league analysts, publish aggregate patterns by season or competition rather than exposing sensitive player-level conclusions without context.

A club building a broader decision-support stack can pair this workflow with cost-effective recruitment platforms for Indian founders for internal hiring, or with automated feedback systems such as AI-based user feedback categorization to organise scout and coach observations. These are adjacent systems, not substitutes for transfer-specific validation.

Limitations and responsible use

Isolation Forest can flag sparse but important groups, favour records with missingness patterns, and confuse changes in reporting practice with market anomalies. Small Indian football datasets also make high-dimensional models unstable. Use fewer, defensible features and report uncertainty.

Protect personal information, respect source licences and avoid inferring sensitive characteristics from nationality, age or other attributes. Do not label a player, agent or club as fraudulent solely because a model produced a high anomaly score. Store data provenance, access controls and correction workflows from the start. If the product will be used by multiple clubs, document how league differences and reporting bias are handled.

Practical checklist

Before deploying the model, confirm that you have:

• A documented data dictionary and source history.
• Separate handling for disclosed, undisclosed, free and loan transfers.
• Peer-group comparisons by season, position and competition.
• Versioned preprocessing, model parameters and random seed.
• A human review queue with reasons and evidence for every flag.
• Monitoring for distribution changes when new seasons or sources arrive.

Isolation Forest is most valuable as a disciplined triage layer. Used with clean data, peer comparisons and expert review, it can help Indian football organisations find errors, surface market inefficiencies and focus limited analyst time on the transfers that deserve a closer look.