- The paper presents a method using machine learning and linear programming to create optimal lineups for Daily Fantasy Football.
- The method trains a neural network on historical data to predict player points and uses linear programming to select optimal lineups within contest rules.
- The method's lineups outperformed random ones but placed in the 31st percentile against human users in real contests, suggesting areas for future model refinement.
The paper entitled "Method and Validation for Optimal Lineup Creation for Daily Fantasy Football Using Machine Learning and Linear Programming" proposes a sophisticated approach for optimizing lineups in Daily Fantasy Sports (DFS) contests, specifically focusing on NFL games. The paper integrates supervised learning through neural networks (NN) with mixed integer linear programming (MILP) to project player performance and maximize fantasy points (FPTS) within the constraints of a DFS contest.
Methodology
- Data Aggregation and Preprocessing:
- Historical player and team performance data from the 2018 NFL season were aggregated. Key features included past FPTS, point differentials, offensive and defensive rankings, game location, latitude and longitude, and betting lines.
- Players were categorized based on position using binary encoding to avoid misleading ordinal implications.
- Neural Network Training:
- A two-layer feed-forward NN was employed, consisting of 43 input features and a single output node predicting FPTS. The architecture included one hidden layer with 19 neurons using sigmoid activation.
- Training employed a Bayesian regularization backpropagation algorithm, using 80% of the data for training and 20% for validation, focusing on minimizing mean squared error (MSE) between predicted and actual FPTS.
- Lineup Optimization:
- An MILP was formulated to select the optimal lineup by maximizing the predicted FPTS while adhering to constraints such as salary cap and positional requirements.
- Optimization was performed iteratively for different configurations of the 'Flex' player position, with the final lineup selected based on maximum predicted FPTS.
- Statistical Validations:
- Robustness of the selection was ensured by simulating 10,000 potential lineups, selecting the most frequently occurring optimal lineup (modal lineup).
- Statistical comparisons were made against randomly generated lineups and real-world user lineups in DFS contests.
Results and Evaluation
- The generated lineups consistently outperformed randomly created counterparts, with average performances above the median, signifying the efficacy of the method.
- In real-world DFS contests across NFL weeks 6 to 16, the generated lineups typically fell in the 31st percentile, outperforming approximately one-third of human users.
- Performance of lineups declined significantly from week 13 onwards, potentially due to factors like increased injury prevalence, playoff considerations, or 'tanking' strategies by NFL teams, which were not accounted for in the NN model.
Discussion and Future Work
- The paper demonstrates that optimized lineups can offer a significant advantage over random selections but show room for improvement against skilled users in DFS contests.
- Future enhancements could include refining the NN by optimizing features and network parameters, incorporating factors such as playoffs dynamics or player injuries/health status.
- Potential applications of the methodology extend beyond NFL DFS contests to other sports or DFS formats by adjusting constraints and prediction models accordingly.
The research provides a baseline framework that integrates advanced machine learning with optimization techniques for strategic lineup creation, establishing a foundation for further refinements and applications in the evolving domain of fantasy sports analytics.