Players’ Performance Prediction for Fantasy Premier League, Using Transformer-based Sentiment Analysis on News and Statistical Data

Sports analysis widely employs data science techniques to derive insights across various sports disciplines (Chmait and Westerbeek, 2021), given the vast amount of data generated during games and training sessions, including both individual and team-based disciplines such as tennis (Giles et al., 2020) and football (Ati et al., 2023). One critical insight derived from sports data is performance analysis, which plays an important role across all sports, particularly in team sports. It serves as an essential tool for managers and coaches (Wright et al., 2014), enabling informed decisions regarding player recruitment, dismissals, tactical adjustments, and game strategies. The ability to forecast player performances and assemble the most effective team for upcoming events is valuable to sports managers (Szymanski and Smith, 1997). This strategic selection process can help a team to win in a competition.

Fantasy Premier League (FPL) is a game platform where people can select their own team from all players in EPL (English Premier League—an English professional league for men’s association football clubs—) and compete with each other in a leader board. FPL stands as the premier fantasy football game worldwide, boasting a record of over 11 million managers participating during the 2023/2024 EPL season (AllAboutFPL 2022; fantasy.premierleague.com). Accurately predicting a high-performing team can significantly benefit FPL players in winning the game. This capability could also be valuable for betting platforms. Additionally, since the prediction process is based on real data from EPL, it can assist team managers and coaches in making informed decisions about player acquisitions, substitutions, and team lineups. Crafting an optimal team within constraints—such as player positions and maximum roster size—transforms the team selection process into a performance prediction and optimization problem. This challenge has been explored in various studies in FPL (Gupta, 2019; Rajesh et al., 2022) as well as other fantasy sports like hockey and basketball (Beal et al., 2020; Hermann and Ntoso, 2015). Our aim in this paper was to provide a novel technique to predict players’ performance for an upcoming event. This method holds potential utility in aiding team selection for FPL competitions.

The EPL dataset is a real-world publicly accessible resource that undergoes weekly updates. It encompasses fundamental statistics for each player, game-week-specific data, and the historical performance of players throughout the season. The majority of methodologies employed in this field leverage this dataset, employing various traditional Machine learning-based (ML-based—“ML-based” is used throughout the paper to refer to all ML methods that do not involve deep learning techniques—) and Deep Learning-based (DL-based) algorithms such as Linear Regression (LR), Decision Trees (DT), Random Forests (RF), Gradient Boosting (GB) Regression, Long Short Term Memory (LSTM), and Convolutional Neural Networks (CNN) (Bangdiwala et al., 2022; Frees et al., 2024; Gupta, 2019; Rajesh et al., 2022) to forecast players’ performance.

Incorporating external sources beyond statistical data for sports analysis has been explored in several studies. Inputs such as match previews (Beal et al., 2021), news articles, and text generated around games have been utilized to enhance the understanding of sporting events and improve the accuracy of performance predictions. For example, Schumaker et al. (2016) demonstrated that analyzing sentiment in tweets could enhance spread and win predictions in the EPL. In FPL domain, research has similarly expanded beyond the confines of statistical datasets to include additional data sources. Bonello et al., (2019), for instance, integrated diverse data streams including betting odds, tweets, and sentiment scores from English blog posts. While they made these datasets available through different APIs (Application Programming Interfaces), most of which were not free, accessing their dataset proved challenging, rendering their methods non-reproducible. In contrast, Baughman et al., (2021) explored a wider array of textual and non-textual data sources, employing techniques such as Document2Vector to incorporate these sources into the prediction process.

In the Premier League, football matches occur weekly, with player data provided in the EPL dataset for each game week. In this paper, we aimed to enhance predictive capabilities by incorporating textual and players’ injury data alongside the EPL dataset. To achieve this, we collected a new dataset comprising news articles about each player and players’ injury for every game week. We then conducted sentiment analysis on the textual data to generate additional features for predicting player performance in the subsequent week and determining their likelihood of participation in the next match. The news dataset was collected from the Google News, The Guardian (open-platform.theguardian.com), and GDELT (blog.gdeltproject.org/gdelt-doc-2-0-api-debuts) APIs, and we employed a pretrained transformer-based sentiment analysis toolkit for labeling sentiment in each article.

We have decided to integrate sentiment of news articles mentioning each player’s name, gathered in the week leading up to each match, to enhance our predictive model. While statistical data provide essential player features for each game, incorporating sentiment of news enriches our dataset by offering insights into the positivity or negativity surrounding each player. Sentiment analysis represents a crucial aspect of Natural Language Processing (NLP), with various tools available that are trained on diverse text corpora for labeling new text as positive, negative, or neutral. In pursuit of this, we have chosen to employ pretrained transformer-based toolkits, for thier advanced capabilities. Transformer structures, renowned for their state-of-the-art design incorporating attention mechanisms (Vaswani et al., 2017), have demonstrated exceptional performance across a spectrum of NLP tasks.

In this paper, we present two methods. In the first method, our predictive tasks involved both regression, for predicting player performance, and classification, to determine if a player will play in the next game week. To address these challenges, we employed various boosting methods, including the GB (Algorithm Friedman, 2001), CatBoost (Dorogush et al., 2018), and XGBoost (Chen and Guestrin, 2016). Alongside news’ sentiment, we extracted approximately 70 features from the EPL dataset.

In the second method, we expanded our dataset to include two sequential seasons using data from The Guardian and GDELT APIs. Additionally, we collected injury data for each player, creating new features for each week before each game. We applied various algorithms, including Ridge Regression (Hoerl and Kennard, 1970), LightGBM (Ke et al., 2017), 1-dimensional CNN (O’shea and Nash, 2015), and LSTM (Hochreiter and Schmidhuber, 1997) architectures. We also compared our model with the state-of-the-art model proposed by Frees et al., (2024). To the best of our knowledge, our approach represents the first instance of employing a pretrained transformer-based sentiment analysis method to extract features from news articles in this domain.

In the upcoming sections of this paper, we will delve into related researches in Section 2. Following this, we will elaborate on our proposed methodologies, encompassing data collection and algorithmic approaches employed for these tasks in Section 3. Subsequently, we will look over the results and evaluations, then in the section 5 we will highlight the value of our work, and discusses its challenges and limitations. Finally, we will draw conclusions, and outline the future research direction in the conclusion section.

Numerous recent studies have focused on performance prediction for FPL, utilizing datasets provided by the FPL API. Additionally, some models have incorporated supplementary data sources, such as tweets or news articles, and metadata on player injuries. These prediction tasks have been explored using both DL-based and ML-based models. A summary of some proposed methods in the literature provided in the Table 1.

Bangdiwala et al., 2022 leveraged three different ML-based models including: LR, DT, and RF. They have considered features such as fixture difficulty, form of the two teams, creativity, and threat of the footballer. Another work provided in 2023 (Shah et al., 2023), considered different rules for team selection and used Multi Criteria Decision Analysis (MCDA) specifically TOPSIS (Technique for Order of Preference by Similarity to Ideal Solution) which consider both the positive and negative aspects of the alternatives in initial team selection. Subsequently, XGBoost Regressor is used to predict the points the team will obtain. Rajesh et al., (2022) employed RF and GB Regression to forecast players’ performance across various positions in a team. In another domain, fantasy basketball team selection tackled by Hermann and Ntoso, (2015), they have used LR model based on box score statistics and modified multinomial naive Bayes classifier to this aim.

Several DL-based models have also been developed in this domain using the EPL dataset, Gupta, (2019) employed a combination of Auto-regressive Integrated Moving Average (ARIMA) with Recurrent Neural Networks (RNNs) and LSTM to predict player points as a team in time series data. Linear Programming was used to optimize the total points by selecting the most promising team based on the calculated predictions. In a research thesis, Ramdas, (2022) introduced a method that combines CNN and LSTM networks to predict players’ performance. Another work in 2024, (Lombu et al., 2024) proposed a method and compared the performance of CNN and LSTM to predicting performances in FPL.

Incorporating other dataset and features rather than the ones provided by FPL API, Bonello et al., (2019) introduced a methodology for team selection in EPL, combining statistical analysis, historical data, and human sentiment. They employed GB Machine to determine player selection, utilizing inputs such as player historical data, FPL API data, betting odds, and sentiment extracted from Twitter and web articles. They accessed diverse datasets from various domains via APIs, although many of these APIs were not accessible for free at the time of this paper’s writing, limiting their utilization. The datasets included betting odds obtained through the football-api (Rapid API), tweets extracted using the Tweepy library, and sentiment scores for English blog posts retrieved for the top 100 search results per query via the Aylien API. These data sources were utilized to analyze player performance in each previous match throughout the current football season. Another novel machine learning system designed to enhance ESPN Fantasy Football team management by analyzing vast amounts of natural language text and multimedia data proposed by researchers from IBM and Disney ESPN (Baughman et al., 2021). The system employs trained statistical entity detectors and DL models to understand and classify player performance trends. The system’s training data includes web archives, ESPN statistics, and injury reports, with 2017 fantasy football data used for testing. Overall, the system aims to address the challenge of information overload in fantasy sports management and improve decision-making processes. This method, utilizes a ML pipeline involving statistical entity detectors and document2vector models to comprehend natural language from over 50,000 sources. DL feed-forward neural networks are then employed to classify players.

The most recent approach in this domain was proposed by Frees et al., (2024). They introduced a novel method based on a 1-dimensional CNN and a transfer learning model that utilized news data about each player collected from The Guardian. Their results showed that the transfer learning model did not perform well, while the CNN architecture outperformed both the transfer learning method and other machine learning-based approaches.

Considering the innovative approaches in the domain of performance prediction for FPL, which include techniques leveraging only the EPL API dataset as well as those incorporating additional data sources, we introduce a novel method. Our approach integrates additional textual data collected for each game week, followed by sentiment analysis using transformer-based techniques.

In this section, we provide an overview of the data utilized in our two methods. Next, we define the classification and regression models employed in this study. Finally, we discuss the evaluation methods applied to assess model performance.

In this section, we examine the impact of incorporating sentiment analysis on news articles and adding players’ injury data as additional features alongside the EPL dataset. We will report the evaluation metrics for all mentioned algorithms utilized in both classification and regression tasks, focusing on how these additional features affect performance.

This study presented a solution for predicting player presence and performance in FPL by augmenting the EPL dataset with external features. The integration of news sentiment and player injury data highlights the importance of leveraging external sources to complement the statistical dataset. Similar studies have shown that incorporating external data can enrich sports prediction models by providing valuable insights into player and team dynamics (Bonello et al., 2019; Baughman et al., 2021, Beal et al., 2021). Our findings demonstrate the potential of incorporating these features to enhance predictive models. Moreover, our approach illustrates the value of leveraging textual data through advanced NLP techniques, such as pretrained transformer-based models. These methods allow for the extraction of insights from unstructured text, which statistical features alone cannot provide. While our approach performed well within the specific context of FPL, its generalizability to other fantasy sports platforms or sports prediction tasks remains untested.

The inclusion of news sentiment improved model performance in the classification task in the first method. Specifically, CatBoost achieved a high performance. These improvements underscore the utility of integrating contextual external data, consistent with findings from Schumaker et al. (2016), who highlighted the importance of sentiment analysis in sports prediction. However, despite these gains, the simplistic baseline model achieved the highest accuracy and F1 score due to its perfect recall. While the baseline captured all participants, it suffered from low precision, indicating a high rate of false positives.

The dataset’s inherent imbalance—where the majority of players do not participate in a given match—may have contributed to the performance gap observed between the baseline and advanced models in the classification task. Incorporating balancing techniques, could improve model reliability. Moreover, the high performance of the baseline raises questions about the added value of more complex algorithms. While boosting models demonstrated better precision by reducing false positives, the baseline’s simplicity and recall advantage highlight the need to evaluate the trade-offs between model complexity and accuracy. A broader range of baselines to contextualize the performance of advanced models could be considered.

In the first method’s regression task, we observed minimal performance differences between models trained with and without news sentiment. One potential explanation is the filtering process, which reduced the dataset size. This reduction may have limited the model’s ability to utilize the added features effectively. Nonetheless, the results suggest that while sentiment features provide benefits for classification tasks, their impact on regression tasks vary depending on the dataset size and the nature of the target variable. To further explore this effect, we implemented the second method using a larger dataset to better assess the influence of the additional features on model performance.

In the second method, the addition of news sentiment and player injury features had mixed effects across algorithms and player positions. For instance, CNN+F demonstrated better performance compared to CNN alone (Frees et al., 2024). Furthermore, the addition of these features enabled CNN+F to achieve the best average performance across all player positions, highlighting the significance of these features in enhancing model performance and surpassing the state-of-the-art model proposed by Frees et al. (2024).

This study faced several challenges, primarily related to data access and dataset size. Data collection posed significant challenges due to API restrictions and access limitations. For instance, the inability to access Twitter data after February 2023 forced us to rely on alternative sources such as Google News, The Guardian, and GDELT APIs. These APIs imposed their own constraints, including request limits and restricted retention periods, complicating the process of gathering consistent and high-quality data especially in the first method using Google News API. These limitations influenced the volume and variety of data available for analysis, potentially constraining the generalizability of our findings.

The limited size of our dataset, particularly in the first method, posed challenges for training both boosting models in regression task and DL models like LSTM. Neural networks are known to require extensive datasets to effectively capture sequential patterns and temporal dependencies (Hochreiter & Schmidhuber, 1997). However, our dataset, covering only one month, was insufficient to train the LSTM models to their full potential, resulting in suboptimal performance. This limitation underscores the importance of obtaining larger and more diverse datasets for future studies.

In conclusion, this study underscores the critical role of incorporating external data sources, such as news sentiment and players’ injury, in enhancing predictive modeling tasks in the domain of FPL. By augmenting the EPL dataset with new additional features, we observed improvements in model performance for both classification and regression tasks. Our findings highlight the importance of diversifying data sources to create more comprehensive datasets, which ultimately contribute to more accurate predictions. In the future, research efforts could be directed towards exploring additional sources of external data and refining predictive models to enhance performance in FPL analysis. These additional resources may include tweets from Twitter, weather data, injury reports, and social media posts authored by each player. As a potential future direction, acquiring a Twitter API account could allow for the collection of more accurate and diverse data from Twitter over several months or years, thereby expanding the dataset. Another research direction could involve further investigation into feature engineering, specifically for LSTM, to better utilize its capabilities in time series problems. Moreover, given the performance of CNN and LSTM, developing a method to leverage a combination of these two models could be beneficial.