Chapter 31 Further Reading: The Complete ML Betting Pipeline
The following annotated bibliography provides resources for deeper exploration of the production ML engineering concepts introduced in Chapter 31. Entries are organized by category and chosen for their relevance to building and operating automated sports betting systems.
Books: ML Engineering and Production Systems
1. Huyen, Chip. Designing Machine Learning Systems. O'Reilly Media, 2022. The single best resource for understanding production ML system design. Covers data engineering, feature engineering, model development, deployment, monitoring, and continual learning. The chapters on data distribution shifts and monitoring are directly applicable to detecting when a betting model's edge is degrading. Huyen's treatment of batch vs. real-time serving architectures maps closely to the pre-game vs. live betting distinction.
2. Burkov, Andriy. Machine Learning Engineering. True Positive Inc., 2020. A practical guide to the engineering side of ML, covering project structure, data pipelines, reproducibility, model serving, and testing. The chapter on testing ML systems provides a framework for the integration tests needed in a betting pipeline. The discussion of model versioning and experiment tracking complements the model registry concept from Chapter 31.
3. Sato, Danilo, Wider, Arif, and Windheuser, Christoph. "Continuous Delivery for Machine Learning." martinfowler.com, 2019. An influential article (expanded into a reference architecture) describing CI/CD practices for ML systems. Introduces the concept of the ML pipeline as a first-class artifact that should be versioned, tested, and deployed alongside the model. The testing pyramid for ML (data tests, model tests, infrastructure tests) is essential reading for building reliable betting pipelines.
4. Kleppmann, Martin. Designing Data-Intensive Applications. O'Reilly Media, 2017. The authoritative reference on data system architecture, covering databases, distributed systems, batch and stream processing. While not ML-specific, the chapters on data models, storage engines, and stream processing provide the foundation for building robust data ingestion and feature store systems. Essential reading for anyone building a pipeline that must be reliable and scalable.
Books: Sports Betting Systems
5. Miller, Ed and Davidow, Matthew. Weighing the Odds in Sports Betting. The Odds Press, 2019. A practical guide to sports betting that covers bankroll management, line shopping, and the mathematics of profitable betting. The Kelly criterion discussion and the analysis of when to bet and when to pass are directly relevant to the execution engine described in Chapter 31. The chapter on sportsbook behavior and bet limits provides essential context for automated execution.
6. Poundstone, William. Fortune's Formula. Hill and Wang, 2005. The history of the Kelly criterion and its application to gambling and investing. Provides the mathematical and historical context for why fractional Kelly is preferred in practice, including the stories of real-world systems that failed due to overbetting. Essential background for understanding the risk management layer of a betting pipeline.
7. Vaughan Williams, Leighton, ed. Information Efficiency in Financial and Betting Markets. Cambridge University Press, 2005. A collection of academic papers examining market efficiency in betting markets. Relevant chapters cover how quickly information is incorporated into odds (critical for understanding the latency requirements of a betting pipeline), the profitability of various betting strategies, and the relationship between bookmaker margin and market efficiency.
Technical Resources: Feature Stores and MLOps
8. Feast Documentation (feast.dev) Feast is the most widely adopted open-source feature store. The documentation provides a practical reference for implementing feature definitions, point-in-time joins (critical for backtesting without leakage), and online/offline serving architectures. The tutorial on building a feature store for a prediction service maps directly to the feature store design in Chapter 31.
9. MLflow Documentation (mlflow.org) MLflow provides experiment tracking, model registry, and model serving capabilities. The model registry component supports the model lifecycle management described in this chapter: registering models with metrics, transitioning between stages (staging, production, archived), and serving models via REST API. The tracking component logs parameters, metrics, and artifacts for every training run.
10. Airflow Documentation (airflow.apache.org) Apache Airflow is the most mature workflow orchestration platform for ML pipelines. The documentation on DAG (Directed Acyclic Graph) design, task dependencies, retries, and alerting is directly applicable to scheduling the batch components of a betting pipeline. The concept of idempotent tasks and backfilling aligns with the data pipeline reliability requirements discussed in this chapter.
Academic Papers and Industry Reports
11. Sculley, D., et al. "Hidden Technical Debt in Machine Learning Systems." Advances in Neural Information Processing Systems (NeurIPS), 2015. The foundational paper on technical debt in ML systems. Identifies the challenges of maintaining production ML: data dependency management, feedback loops, pipeline jungles, and configuration debt. Every issue raised in this paper manifests in a betting pipeline. The concept of "changing anything changes everything" is particularly relevant when feature definitions or data sources are updated.
12. Breck, Eric, et al. "The ML Test Score: A Rubric for ML Production Readiness and Technical Debt Reduction." IEEE International Conference on Big Data, 2017. Proposes a scoring system for evaluating the production readiness of ML systems across four dimensions: data tests, model tests, ML infrastructure tests, and monitoring tests. This rubric can be directly applied to evaluate a betting pipeline's readiness for live deployment. A pipeline that scores poorly should not be trusted with real capital.
13. Paleyes, Andrei, Urma, Raoul-Gabriel, and Lawrence, Neil D. "Challenges in Deploying Machine Learning: A Survey of Case Studies." ACM Computing Surveys, 55(6), 2022. A comprehensive survey of real-world ML deployment challenges across industries. The identified challenges --- data management, model selection, deployment, monitoring, and organizational --- are all present in betting pipeline development. The case studies provide cautionary tales and best practices applicable to the sports betting context.
Technical Resources: APIs and Execution
14. FastAPI Documentation (fastapi.tiangolo.com) FastAPI is the recommended framework for building the prediction serving API described in this chapter. The documentation covers async request handling, automatic data validation with Pydantic, dependency injection, and automatic API documentation. The performance characteristics (async, high throughput) make it suitable for both batch and real-time prediction serving.
15. The Odds API Documentation (the-odds-api.com) A practical resource for accessing real-time odds data from multiple sportsbooks via a single API. The documentation demonstrates the data structures and rate limiting patterns used in the odds ingestion component of a betting pipeline. Understanding the available endpoints, update frequencies, and data format is essential for building the data layer.
16. Redis Documentation (redis.io) Redis serves multiple roles in a betting pipeline: as a cache for frequently accessed features (reducing feature store latency), as a message broker between services (using Redis Streams), and as a session store for tracking active bets. The documentation on data structures, persistence, and pub/sub patterns provides the building blocks for low-latency pipeline components.
Monitoring and Observability
17. Prometheus and Grafana Documentation (prometheus.io, grafana.com) Prometheus (metrics collection) and Grafana (visualization and alerting) form the standard monitoring stack for production systems. The documentation covers metric types (counters, gauges, histograms), alerting rules, and dashboard design. For a betting pipeline, custom metrics include prediction latency, feature freshness, model calibration error, and daily P&L.
18. Shankar, Shreya, et al. "Operationalizing Machine Learning: An Interview Study." arXiv:2209.09125, 2022. An interview study with ML engineers at major tech companies about the challenges of operating ML systems in production. Key findings include the prevalence of data quality issues as the primary source of model failures, the importance of monitoring beyond just accuracy metrics, and the difficulty of debugging distributed ML pipelines. These findings directly inform the monitoring and alerting architecture described in Chapter 31.
How to Use This Reading List
For readers working through this textbook sequentially, the following prioritization is suggested:
- Start with: Huyen (entry 1) for a comprehensive ML systems foundation, and Feast documentation (entry 8) for hands-on feature store practice.
- Go deeper on reliability: Sculley et al. (entry 11) and Breck et al. (entry 12) for understanding technical debt and production readiness.
- Go deeper on execution: Miller and Davidow (entry 5) for sports betting mechanics, and Poundstone (entry 6) for Kelly criterion context.
- For implementation: FastAPI (entry 14) and MLflow (entry 9) for building the serving and registry components.
Many of these resources will be referenced again in later chapters as pipeline concepts are applied to NLP integration, live betting systems, and multi-sport operations.