Sports betting integrity & match-fixing detection

Match-fixing is usually organised around betting: someone influences a result or an event within a match and profits on bets. Because the money flows through betting markets, those markets leave traces: odds that move without public news, unusual volumes, in-play bets that anticipate events. AI models monitor odds across many bookmakers and compare them with the expected market behaviour for the match. They flag outliers for human integrity analysts, whose reports feed sporting sanctions and criminal cases. The model produces a lead, never a verdict.

Problem & value

• Who: sports federations and leagues (e.g. football associations, tennis and basketball bodies), national gambling regulators, licensed betting operators and, increasingly, prediction-market operators. SMB-sized clubs and lower leagues are the weakest link because they have little oversight and poorly paid players, but they rarely build tooling themselves.
• Today: federations buy monitoring services, run hotlines and education, and rely on information exchange with operators and police. Analysts review alerts manually, using odds histories and context such as injuries or weather.
• Why AI helps: the volume is far beyond manual review. Sportradar's integrity page reports more than 1,000,000 events monitored per year, 11,000+ suspicious matches detected, 70+ sports, and support for 1,000+ sporting sanctions and 70+ criminal convictions [1]. These are vendor-reported, undated figures; treat them as indicative. Wikipedia cites the same vendor for 881 suspicious football matches in 83 countries in 2023 and 721 in 72 countries in 2024 [2]; those are secondary figures and not independently verified here.
• Caveat: "suspicious" means an unexplained betting pattern, not proven manipulation. Many alerts have innocent explanations (late team news, a sharp bettor, a data error).

Variants

• Federation/league service: monitor all competitions under a federation's jurisdiction, with sanction support and prevention programmes. UEFA and the DFL are named Sportradar partners [3].
• Operator-side integrity: a bookmaker monitors its own markets and accounts, reports to regulators and to information-sharing bodies, and limits stakes on risky events.
• Regulator-side national platform: the Macolin Convention (Council of Europe, concluded 2014, in force since 1 September 2019) targets manipulation of sports competitions and illegal betting; it envisages cooperation between sport, operators and authorities [4]. Check the current ratification status per country; the source consulted lists 15 parties and does not show Germany among them.
• Individual-sport and in-play (spot-fixing) monitoring: tennis, table tennis and similar sports with many low-level events; models focus on in-play markets and single-point events.
• Prediction markets and exchanges: event contracts on sports add new data sources and new manipulation channels; Sportradar announced a Polymarket partnership covering 20+ leagues in 2026 [3].
• Esports and e-gaming integrity: same method, different feeds; for cheating in the game itself see ANOM-036.
• Out of scope here: securities market abuse (ANOM-012), generic AML (ANOM-002), payment fraud (ANOM-001).

How it works

1. odds feeds (many books) + official match data + reference data
2. normalisation & time alignment (streaming)
3. market-model / expected-odds baseline per sport, league and market type
4. anomaly scoring (odds drift, volume spikes, leader-follower timing, in-play deviations)
5. ranking & deduplication of alerts
6. analyst review (human) with LLM-drafted summary and evidence timeline
7. report to federation / regulator / police via the legally defined channel
8. feedback of outcomes into labels and model retraining

• Baseline models: gradient-boosted or neural models learn what the market should look like given the teams, league and time to kickoff; large, unexplained deviations are scored. Time-series and sequence models compare movements across bookmakers to detect which markets moved first.
• Graph analytics: link accounts, devices, stakes and timing at operator level (only where lawful) to identify coordinated betting groups; same technique as in ANOM-022 and ANOM-023.
• Unsupervised plus labelled: confirmed cases are few, so scoring combines anomaly detection with supervised models trained on sanctioned cases. Label scarcity and survivorship bias (undiscovered fixes look like normal matches) are structural.
• LLM layer: summarise alert context, retrieve news such as injuries and line-ups, and draft the analyst report. The LLM never decides whether manipulation occurred.
• Human-in-the-loop: integrity analysts triage; a disciplinary panel or prosecutor decides. Players are never sanctioned on a score alone.

Data preparation

Integrity detection depends on high-volume, time-stamped data: odds and market-volume feeds from many bookmakers, official match events, fixture and player reference data, account-level betting data where legally shared, and intelligence reports or hotline tips. In practice this lives in vendor feeds, operator data warehouses, case-management systems and mail or document archives of integrity units.

Typical gaps: inconsistent identifiers for matches, players and markets across providers; missing timestamps or time zones; gaps in lower-league coverage; unlabelled historic cases; account data that may not be shared across operators; and case files that exist only as scans or e-mails.

Preparation steps:

1. Map sources and gaps against the use case (PREP-002) and set up connectors for feeds and warehouses (PREP-005).
2. Build an entity master (match, team, player, market IDs) and a shared glossary (PREP-012).
3. Harvest labels from past confirmed and cleared cases (PREP-020) and build an expert-validated test set (PREP-018).
4. Digitise paper or scanned case files (PREP-006).
5. Pseudonymise account data and clarify the legal basis (PREP-022, PREP-025).

Tools: ETL pipelines, a time-series database, OCR for case files. Rough share of effort: 50 to 60 percent of the project, as data harmonisation dominates.

Implementation paths

Which path for whom

• enterprise: A. Operators and federations buy established integrity-monitoring services with broad odds coverage; building that data network yourself is not realistic.
• developer: C. Open-source anomaly-detection and streaming stacks on own hardware give full control over models and data, with odds feeds licensed separately.
• public: B. Regulators and prosecutors build case tooling on sovereign EU cloud APIs, combining licensed feeds with their own data under procurement and sovereignty rules.

Path A — Buy (off-the-shelf)

Integrity-monitoring services are the normal route: Sportradar Integrity Services (UFDS AI, Swiss/Norwegian-rooted, listed company) [1], Genius Sports (used by MLB for monitoring betting on games [2]), and the International Betting Integrity Association (IBIA), a not-for-profit alert-sharing body of licensed operators. Federations also use the IOC's Integrity Betting Intelligence System and the tennis integrity unit's sector-specific tools; check each body's current offer on its own site, as I could not re-verify them in this research round. EU/European: Sportradar's integrity business is Europe-based (St. Gallen); check data-processing location in the contract. Right choice for virtually every federation, regulator and operator: the value is in cross-bookmaker data coverage, which nobody can replicate alone.

Path B — Build on cloud APIs

Only sensible for the analyst-assistance layer: a European-hosted LLM (e.g. Mistral via La Plateforme, or Azure OpenAI/AWS Bedrock in EU regions) to summarise alerts, search news and draft reports, orchestrated with an agent framework and the case-management system. The scoring itself needs licensed odds data (Sportradar, Genius Sports, Betfair exchange data, bookmaker APIs) and your own models in a cloud data platform. Integrate with existing case tools and send results by secure channel to authorities.

Path C — Self-host (open source)

Feasible for a national regulator or large operator that has data access: Kafka or Flink for streaming, scikit-learn, LightGBM/XGBoost and PyTorch for models, PyOD/Isolation Forest-type detectors, NetworkX or Neo4j for linkage, and an open-weight LLM (e.g. Mistral, Qwen, Llama class, 7–30B) for report drafting. CPU is enough for scoring; HW1 suffices, HW2 helps for the LLM. The binding constraint is data licences and legal basis, not compute.

Path D — Custom training

Worth it for operators with many years of own market data and a case archive. Train sport- and league-specific baselines and supervised classifiers on sanctioned cases; use time-split validation to avoid leakage. Pool labels with sector bodies where legally possible. Expect few positive labels (hundreds, not millions).

State of the art (2026)

• Commercial systems combine continuous cross-bookmaker odds monitoring with machine learning and expert analysts; the human-analyst step remains central [1].
• Trend: expansion from football to many sports and to prediction markets/exchanges, where event contracts add new data and manipulation incentives [3]. Growing cooperation between operators, federations and regulators under national platforms.
• Recent work focuses on in-play and micro-event markets (spot-fixing), where signals are short-lived, and on graph-based account linkage.
• Limits and failure modes: false positives from legitimate information advantages; small-liquidity markets (lower leagues) that are noisy and easy to move; undetected fixing without betting traces (sporting-motive fixing); adversaries adapting stake patterns; data licence cost; very few labels. Public benchmark datasets of confirmed cases do not exist, so published accuracy figures are vendor-specific.
• Notable historical cases show the stakes: the 2005 German Bundesliga referee scandal, a 2013 Estonian case with 11 men charged over 17 matches, and a 2023 tennis ring involving at least 181 players [2].

Shared resources & combinations

• Shared building blocks: streaming ingestion, anomaly-detection and tabular-ML stack, graph analytics for linkage, an LLM case-summarisation layer, and case-management with audit trail. All run on CPU-class hardware (HW1) with an optional small LLM.
• Combines with: ANOM-012 (market abuse; same surveillance architecture and analyst workflow), ANOM-002 and ANOM-023 (money-flow tracing when fixing proceeds are laundered), ANOM-022 (mule accounts used for stakes), ANOM-004 (fake and multiple betting accounts), ANOM-036 (esports/online game cheating), ANOM-001 (payment fraud at operators).
• Shared data: odds feeds and official data feeds are expensive; one licence can serve integrity, risk management and trading teams within an operator.

Effort, cost & KPIs

• Path A: weeks to onboard, annual service fee dominated by competition coverage; internal analysts still needed (roughly 1–3 FTE for a large federation, an estimate that depends on scope).
• Path B/C/D: quarters; main costs are data licences, data engineering, labelling and analyst time. Models are cheap compared with data.
• KPIs: share of alerts confirmed by analysts (precision), alerts per 1,000 matches monitored, time from market move to alert, number of cases leading to sanctions or prosecutions, coverage of competitions, and rate of alerts overturned later. Track false positives separately for lower leagues.

People & change

Who and when. Management and the integrity or compliance lead sponsor the project. Data owners of odds feeds, account data and case archives are involved first, because access and data-sharing agreements decide feasibility. IT security protects sensitive case data; legal/compliance covers gambling law, cooperation with regulators and evidence rules. Suppliers (data providers) and the DPO join before the pilot.

Works council. No in the typical setup, as the system analyses markets, accounts and matches rather than employee behaviour or performance. If analyst activity is logged in a way that allows performance evaluation, §87(1) no. 6 BetrVG can still apply, and §90(1) no. 3 calls for timely information when work processes change. Public bodies involve the Personalrat in that case.

Data protection. Account-level betting data is personal data; a DPIA (Art. 35 GDPR) is likely because individuals are systematically evaluated and may face sanctions. Use AVV (Art. 28 GDPR) with vendors and involve the DPO early (GUIDE-005).

AI literacy and control. Train analysts on score meaning, base rates and false positives (GUIDE-006). Alerts are leads, never verdicts: a human analyst reviews every case before any report to a federation or authority. Start with a small analyst pilot group, measure precision and document decisions (GUIDE-009).

Risks, regulation & compliance

• EU AI Act: monitoring odds data about matches is not itself a listed high-risk use. Classification is depends: if a law-enforcement authority uses the output to assess the reliability of evidence or to profile natural persons, Annex III point 6 (law enforcement) may apply. Art. 5(1)(d) prohibits risk assessments that predict that a person will commit a criminal offence based solely on profiling or personality traits; do not build player or bettor "propensity to fix" scores from profiling alone. Keep outputs match-level and evidence-based, with a human decision on each case. Verify against the Official Journal text [5].
• GDPR: match-level odds are not personal data, but operator account data, device identifiers and player names are. Criminal-offence data (Art. 10) needs a legal basis in national law. Information sharing between operators, federations and authorities needs a clear statutory basis; the Macolin Convention encourages national platforms [4]. Do a DPIA and minimise data.
• Sector law (Germany, to be confirmed with counsel): sports-betting fraud and manipulation of professional sporting competitions are criminal offences in the German Criminal Code (§§ 265c, 265d StGB); gambling supervision sits with the Länder authority GGL under the Glücksspielstaatsvertrag. Other states differ.
• Risks: wrongly accusing athletes (reputational and employment harm), data-licence breaches, bias against lower-income leagues, leakage of investigation details, and adversarial adaptation. Mitigations: analyst confirmation, documented evidence chain, appeal rights for sanctioned persons, strict access control and audit logging.
• Not legal advice.

Getting started

1. Define your role (federation, operator, regulator) and the legal channel for reporting and data exchange; clarify the GDPR basis first.
2. Map competitions at risk (lower leagues, youth, individual sports with in-play markets) and request quotes from monitoring providers.
3. Set up an alert-handling process with named analysts, escalation rules and a case file template.
4. Add prevention: education for players and officials, and an anonymous reporting hotline.
5. Pilot an LLM assistant for summaries on past, anonymised alerts; measure analyst time saved before expanding.

Sources

1. Sportradar, Integrity Services overview (vendor figures, undated): https://sportradar.com/integrity-services/ (accessed 2026-10)
2. Wikipedia, Match fixing (detection approaches, cases, vendor-sourced figures): https://en.wikipedia.org/wiki/Match_fixing (accessed 2026-10)
3. Sportradar corporate site (clients/partners, 2026 news): https://www.sportradar.com/integrity-services/universal-fraud-detection-system/ (accessed 2026-10)
4. Wikipedia, Macolin Convention (adoption 2014, in force 2019, parties): https://en.wikipedia.org/wiki/Macolin_Convention ; official Council of Europe page: https://www.coe.int/en/web/sport/macolin-convention (not retrievable in this session; confirm on the official page)
5. Regulation (EU) 2024/1689 (AI Act), Official Journal: https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202401689 (article references from the regulation; text not retrievable in this session, verify before relying)