Sports Analytics Weekly by kubeia.io - 40/2023

Abstract:The paper underscores how decentralization in sports betting addresses the drawbacks of traditional centralized platforms, ensuring transparency, security, and lower fees. Non-custodial solutions empower bettors with ownership of funds, bypassing geographical restrictions. Decentralized platforms enhance security, privacy, and democratic decision-making. However, decentralized sports betting necessitates automated market makers (AMMs) for efficient liquidity provision. Existing AMMs like Uniswap lack alignment with fair odds, creating risks for liquidity providers. To mitigate this, the paper introduces UBET AMM (UAMM), utilizing smart contracts and algorithms to price sports odds fairly. It establishes an on-chain betting framework, detailing market creation, UAMM application, collateral liquidity pools, and experiments that exhibit positive outcomes. UAMM enhances decentralized sports betting by ensuring liquidity, decentralized pricing, and global accessibility, promoting trustless and efficient betting.

Soccer’s biggest event will celebrate its centenary by placing games in South America, Europe and Africa. The decision could pave the way for Saudi Arabia to host in 2034.

Games in the United States are dictated by broadcast schedules. In most countries a 3 PM kick off means the players are on the field at 2:55 PM and they kick off at exactly 3 PM with the referee dictating the start time. In the US, where pregame shows aren’t that common, the program would begin at 3 PM as the players enter the field and observe the national anthem, with kick-off at 3:08 PM with the ‘red hat’ from the broadcaster signalling to the referee that he can start the game.

Abstract:The inherently diverse and uncertain nature of trajectories presents a formidable challenge in accurately modeling them. Motion prediction systems must effectively learn spatial and temporal information from the past to forecast the future trajectories of the agent. Many existing methods learn temporal motion via separate components within stacked models to capture temporal features. This paper introduces a novel framework, called Temporal Waypoint Dropping (TWD), that promotes explicit temporal learning through the waypoint dropping technique. Learning through waypoint dropping can compel the model to improve its understanding of temporal correlations among agents, thus leading to a significant enhancement in trajectory prediction. Trajectory prediction methods often operate under the assumption that observed trajectory waypoint sequences are complete, disregarding real-world scenarios where missing values may occur, which can influence their performance. Moreover, these models frequently exhibit a bias towards particular waypoint sequences when making predictions. Our TWD is capable of effectively addressing these issues. It incorporates stochastic and fixed processes that regularize projected past trajectories by strategically dropping waypoints based on temporal sequences. Through extensive experiments, we demonstrate the effectiveness of TWD in forcing the model to learn complex temporal correlations among agents. Our approach can complement existing trajectory prediction methods to enhance prediction accuracy. We also evaluate our proposed method across three datasets: NBA Sports VU, ETH-UCY, and TrajNet .

Abstract:Daily fantasy sports (DFS) are weekly or daily online contests where real-game performances of individual players are converted to fantasy points (FPTS). Users select players for their lineup to maximize their FPTS within a set player salary cap. This paper focuses on (1) the development of a method to forecast NFL player performance under uncertainty and (2) determining an optimal lineup to maximize FPTS under a set salary limit. A supervised learning neural network was created and used to project FPTS based on past player performance (2018 NFL regular season for this work) prior to the upcoming week. These projected FPTS were used in a mixed integer linear program to find the optimal lineup. The performance of resultant lineups was compared to randomly-created lineups. On average, the optimal lineups outperformed the random lineups. The generated lineups were then compared to real-world lineups from users on DraftKings. The generated lineups generally fell in approximately the 31st percentile (median). The FPTS methods and predictions presented here can be further improved using this study as a baseline comparison.

Abstract:This study explores strategic considerations in professional golf's Match Play format, challenging the conventional focus on individual performance. Leveraging PGA Tour data, we investigate the impact of factoring in an adversary's strategy. Our findings suggest that while slight strategy adjustments can be advantageous in specific scenarios, the overall benefit of considering an opponent's strategy remains modest. This confirms the common wisdom in golf, reinforcing the recommendation to adhere to optimal stroke-play strategies due to challenges in obtaining precise opponent statistics. We believe that the methodology employed here could offer valuable insights into whether opponents' performances should also be considered in other two-player or team sports, such as tennis, darts, soccer, volleyball, etc. We hope that this research will pave the way for new avenues of study in these areas.

Abstract:Even though Google Research Football (GRF) was initially benchmarked and studied as a single-agent environment in its original paper, recent years have witnessed an increasing focus on its multi-agent nature by researchers utilizing it as a testbed for Multi-Agent Reinforcement Learning (MARL). However, the absence of standardized environment settings and unified evaluation metrics for multi-agent scenarios hampers the consistent understanding of various studies. Furthermore, the challenging 5-vs-5 and 11-vs-11 full-game scenarios have received limited thorough examination due to their substantial training complexities. To address these gaps, this paper extends the original environment by not only standardizing the environment settings and benchmarking cooperative learning algorithms across different scenarios, including the most challenging full-game scenarios, but also by discussing approaches to enhance football AI from diverse perspectives and introducing related research tools. Specifically, we provide a distributed and asynchronous population-based self-play framework with diverse pre-trained policies for faster training, two football-specific analytical tools for deeper investigation, and an online leaderboard for broader evaluation. The overall expectation of this work is to advance the study of Multi-Agent Reinforcement Learning on Google Research Football environment, with the ultimate goal of benefiting real-world sports beyond virtual games.

The burgeoning field of sports analytics has led to the proliferation of tools that deliver real-time in- sights to coaching staff, aiding them in strategic decision-making during games. However, existing sys- tems focus solely on in-game data, thereby overlooking the benefits of incorporating historical data for deeper, contextual insights. In this paper, we introduce StratAlign, a novel system designed to mine large-scale historical events data to identify similar ball movement patterns, or "trajectories," in football games. We propose a dynamic time warping (DTW)-based distance function that offers robust trajectory comparison, and employ a clustering mechanism to efficiently prune the search space. Furthermore, StratAlign is built with scalability in mind, using a disk-based hash index to maintain the dataset and sev- eral memory-eviction strategies to operate within limited resource constraints. We also address the challenge of aligning event timestamps with actual video footage through an innovative computer vision approach. Our experimental evaluations confirm the system's ability to retrieve relevant ball trajectories in significantly less time compared to an in-memory baseline solution, making StratAlign an efficient and cost-effective tool for real-time strategic analysis in football games.

The collective behaviour of opposing multi-agent teams has been extensively researched in game theory, robotics, and sports analytics. In sports, team tactics frequently

encompass individuals’ strategic spatial and action behaviours, and can be manifested in sequences of events during periods of possession in football. The analysis of team tactics

is critical for training, strategy, and ultimately team success. While conventional notational and statistical analysis approaches can provide valuable insights into team tactics,

contextual information has generally been overlooked, and teams' performance has not been holistically evaluated. To consider contextual information, we employed the

sequential pattern mining algorithm PrefixSpan to extract team tactics from possessions, the Neural Marked Spatio Temporal Point Process (NMSTPP) model to model expected

team behaviour for a fair comparison between teams, and the Holistic Possession Utilisation Score (HPUS) metrics to evaluate teams’ possessions. In the experiments, we

identified five team tactics, validated the NMSTPP model when StatsBomb 360 data was incorporated, and analysed the English Premier League (EPL) teams in the 2022/2023

season. The results were visualised using radar plots and scatter plots with mean shift clustering.

The main contributions of this work are the following:

1. We show how the SoccerMap architecture can be modified to predict passing tendencies in a given situational context.

2. We present an encoder-decoder architecture that can generate a compact embedding for describing how a player’s pass selection decision was affected by

the situational context.

3. We show how these embeddings can be used to get a better insight into a team’s playing style and how they tend to vary their game plan based on the situational

context.

Computing the optimal defensive player positioning in football is challenging but valuable for the decision-making of both players and coaches. Previous studies have utilized

mathematical-based probabilistic models to represent off-ball scoring opportunities (OBSO). However, these studies did not focus on defending sides, where the usage of the

event label is limited. In addition, the previous studies on defensive evaluation cannot explain where each defender should be positioned at that time and evaluate defensive

positioning. In this study, we propose an evaluation method of team defense positioning by computing counterfactuals using StatsBomb 360 data. We also identify the optimal

positioning of the defensive teams by searching counterfactual positionings. This study will quantitatively allow us to evaluate team defenses and help the players and their

coaches more easily suggest ways to improve their decision-making abilities in future games.

This study seeks to broaden the comprehensive understanding of team tactics through an exploration of playing styles and player connectivity during offensive sequences. Specifically,

the study aims to identify key players with substantial involvement and importance in possession sequences. The outcomes of this research will illuminate the key positions within a

team, reflecting the distinct players and their interactions during offensive maneuvers. The central hypothesis posits that analyzing player interactions within the possession framework will

reveal essential connections and interaction patterns among players, leading to a deeper understanding of key attacking positions and team playing styles, in contrast to existing

frameworks, which predominantly focus on ball and player parameters.

Models to value actions within football have existed for over a decade and many people have created different versions within this space, all making different design decisions.

The football analytics community has developed many different models to try and value the contribution of players to their contracted club.

From a club perspective, this has been pursued in order to maximise the footballing output of their financial outlay by finding players who are cheaper than they should be or

who will accept lower wages.

This paper details further work on a novel machine learning framework for forecasting football results building from fundamental actions upwards to predict games and seasons.

The year 2023 has witnessed significant progress in Generative AI. One would have, in fact, needed to have been living under a rock not to notice the surge in AI-based digital

assistants powered by Large Language Models (LLMs). While these models have garnered considerable attention from the public, they have also achieved notable

benchmarks in performance. They have proven their potential in various applications, especially natural language processing (NLP) and sequential data generation, including

program code and protein sequences. However, their potential in the realm of sports remains largely untapped. Thus, in this study, we take the first steps toward uncovering

the potential of state-of-the-art LLMs as tactical analysts by introducing TacticalGPT, an AI-based assistant coach for professional football.

The end goal of this paper is to present a process-aware analysis of event data to discover team playing strategy. We use Process Discovery techniques to retrieve and analyze the inherent patterns of play out of event traces extracted from event data. We present a purpose-driven methodology to reduce the variance and lack of structure in the traces, highlight frequent patterns of play, and facilitate the control flow identification to produce the most accurate models of team strategy. Hence, we contribute to defining a knowledge discovery pipeline to manage and analyze the large-scale availability of event data for team strategy identification. We also demonstrate how the models could be integrated into football-specific visualizations to provide a novel and better understanding of a team's execution and performance during a game. The methodology is evaluated in the 2021/2022 season of the English Premier League.

Men’s and women’s football are different on both strategic and tactical levels - from player physiology and injury to league structure - to the point that models can easily distinguish between mens’ and womens’ games from match variables. Men’s football analytics is expanding and hugely profitable, and models and insights from men’s football are (potentially inappropriately) applied to the women's game. Using the analysis carried out by Soccerment to cluster male players based on their functions as a framework, the aim of this paper is to specifically investigate player functions that can be identified in the women’s game. We use Statsbomb event data from the 2018/19 to 2022/23 Women’s Super League (WSL) to search for correlations between on-ball events and cluster the data based on the function of a player

In recent years, StatsBomb has collected tactical event data (a.k.a. StatsBomb 360 data), which includes the positions of all the visible players when an event happens. With this information provided, it becomes possible to distinguish the different situations from the events at the same location. This work aims to fill this gap. Specifically, our goal is to cluster the freeze frames in StatsBomb 360 data into various situations and conduct performance analysis based on the discovered situations. The original data format of the freeze frames is too complex to apply traditional clustering algorithms directly. To address this issue, we propose first compressing the frames into low-dimensional features via deep representation learning, then clustering them in the feature space to discover different game situations in each pitch area. Furthermore, we propose a novel performance evaluation metric, Situational Expected Threat (Situational xT), which replaces the positional states in the Expected Threat model with our discovered game situations. By evaluating a team’s actions using Situational xT and summarizing their values according to the event situation index, our method can demonstrate in what situations they play more effectively.

This write up is my notes on a few papers looking at using order book data to model price movements.

In this article, we will explore the concept of change point detection in time series data using Python.Change point detection is a powerful technique that helps you identify significant shifts in your time series data, which can provide valuable insights for decision-making and forecasting.However, detecting change points can be challenging, especially when working with noisy or complex data.Several algorithms are available, but choosing the right one and fine-tuning its parameters can be time-consuming and confusing.This is a practical guide on how to apply these algorithms using the Python library Ruptures and real-world data from Google Search Console.This guide is designed for data analysts, data scientists, and anyone interested in working with time series data.

We are excited to announce the release of XGBoost 2.0. This note will begin by covering some overall changes and then highlight specific updates to the package.

Abstract:In this paper, we introduce TimeGPT, the first foundation model for time series, capable of generating accurate predictions for diverse datasets not seen during training. We evaluate our pre-trained model against established statistical, machine learning, and deep learning methods, demonstrating that TimeGPT zero-shot inference excels in performance, efficiency, and simplicity. Our study provides compelling evidence that insights from other domains of artificial intelligence can be effectively applied to time series analysis. We conclude that large-scale time series models offer an exciting opportunity to democratize access to precise predictions and reduce uncertainty by leveraging the capabilities of contemporary advancements in deep learning.