1 Introduction

The global investment environment is experiencing an unprecedented surge in the allocation of savings into a diverse array of mutual funds, accompanied by a steady rise in the number of investors participating in these funds. This phenomenon carries profound implications for the world economy on two distinct fronts. Firstly, mutual funds play an indispensable role as a vital source of financing for corporations across various sectors, facilitating capital flow to fuel business operations, expansion, and innovation. Secondly, these funds serve as the primary conduit through which millions of investors worldwide channel their savings towards achieving long-term financial objectives, be it retirement planning, wealth accumulation, or the pursuit of specific financial goals.

In this context, socially responsible investors are confronted with a unique confluence of factors that shape their investment decisions. The quest for both financial prosperity and the advancement of personal values calls for a discerning approach to mutual fund selection. Beyond the conventional considerations of investment strategies and historical performance, profit-seeking socially responsible investors recognize the essential role that Environmental, Social, and Governance (ESG) ratings play in the contemporary investment landscape. The integration of ESG criteria into their investment assessment represents a conscientious endeavor to harmonize profit motives with ethical principles, aligning their investments with a broader commitment to sustainability and responsible corporate behavior.

Acknowledging the growing significance of ESG factors, mutual fund databases have taken substantial strides in catering to the evolving needs of investors. Leading financial database providers such as Refinitiv Lipper and Morningstar have incorporated ESG ratings into their comprehensive datasets. This proactive measure empowers both financial advisors and individual investors to infuse the ESG dimension into their fund selection process. The availability of accurate, standardized, and comparable environmental, social, and governance data equips investors with a powerful tool to evaluate the social and ethical impact of their investments, transcending mere financial considerations. This marks a transformative shift, ushering in an era where investments are not only expected to yield financial returns but also to contribute positively to a sustainable and equitable future.

In sum, the burgeoning prominence of mutual funds in the global investment landscape necessitates a holistic approach to investment decision-making. For socially responsible investors, the integration of ESG ratings into their investment strategy is emblematic of a commitment to driving positive change while pursuing financial objectives. The accessibility of ESG data within mutual fund databases empowers investors to make informed choices, thereby reinforcing the symbiotic relationship between finance and social responsibility. This confluence of forces underscores the transformative potential of responsible investing and its capacity to shape a more sustainable and ethical future for global financial markets. Morningstar started rating mutual funds according to ESG criteria in 2016 as a response to investors’ growing concern for sustainability values. Morningstar Sustainability Ratings are determined using bottom-up assessments of the underlying holdings within a portfolio, supported by Sustainalytics. The Morningstar sustainability rating is in the form of five different categories (1 to 5 globes): low equals 1 globe and high equals 5 globes. Refinitiv Lipper fund ESG numerical scores provide investors and financial advisors with overall fund-level scores that describe the sustainability of a particular mutual fund. All Refinitiv Lipper fund ESG scores are rolled up from the security level to give a fund-level score.

Table 1 Global sustainable funds (2021–2022)
Full size table

According to the Global Sustainable Fund Flows Report (Morningstar, 2021, 2022) in the global market of sustainable funds, comprising both open-end and exchange-traded funds with a clear focus on sustainability, Europe is leading the sustainable funds market with 83% of global sustainable fund assets, followed by the US with 11% of global sustainable fund assets as of December 2022 (see Table 1). Regarding the number of sustainable funds, Europe also leads the sustainable fund market segment with 5,316 funds (76%), followed by the US with 598 funds (9%) and Asia (excluding Japan) with 459 funds (6%). Although the European fund market plays a prominent role in the sustainable fund segment, US households hold a comparatively larger portion of their assets (about 23%) in regulated mutual funds, whereas European households just hold about 9% of their assets in regulated mutual funds (International Investment Funds Association, 2021). The outstanding portion of US household financial wealth invested in funds and the fact that the US has the largest share of total net assets of worldwide regulated open-end mutual funds are two compelling reasons to focus our research on the US mutual funds market.

Investors’ motivations for opting for mutual funds with robust ESG scores are inherently diverse, encompassing a spectrum of orientations ranging from socially driven to profit-seeking objectives. Recognizing this divergence in investor intent is fundamental, as it underscores the significance of aligning individual motivations with mutual fund choices. Socially responsible investors, characterized by a profound commitment to ESG criteria, often exhibit a willingness to forego a portion of their financial returns in pursuit of the noble cause of promoting sustainable corporate practices. While this apparent trade-off between profit and sustainability might be perceived as a potential drawback, a closer examination might reveal a more nuanced dynamic at play.

Companies that operate under a mandate of social responsibility tend to be characterized by a heightened level of transparency in their operations. This heightened transparency has cascading effects, potentially enhancing their financial performance. Transparency serves as a conduit through which companies can communicate their commitment to ethical conduct and sustainability practices to stakeholders. In doing so, these companies foster trust and confidence among investors, customers, and the broader market, consequently bolstering their overall performance. This suggests that, contrary to the notion of profit sacrifice, investments in companies with strong ESG values may yield not only ethical satisfaction but also competitive financial returns.

Nevertheless, it is imperative to acknowledge the intricate and multifaceted nature of the relationship between social metrics and financial performance. This relationship is contingent upon an array of factors, including prevailing market conditions, the specific investment strategies employed, and the precise ESG criteria employed to evaluate mutual funds. The interplay of these variables results in a complex, context-dependent landscape where the impact of ESG metrics on fund performance exhibits variability. Consequently, it becomes clear that there is a pressing need for comprehensive studies, such as the one presented in this paper, aimed at unraveling the intricate dynamics between ESG metrics and fund performance.

This study seeks to contribute to the ongoing debate by shedding light on the interrelationships between ESG considerations and mutual fund performance. We aim to provide investors, financial advisors, and market participants with valuable insights into the implications of ESG integration within the investment landscape. In so doing, we intend to foster a deeper understanding of how ethical and financial objectives can coexist harmoniously, ultimately driving sustainable and responsible investment practices in the evolving financial markets.

In our study, we focus on the US mutual fund market because it held about 52% of the world’s regulated mutual fund assets in 2021. Within the US mutual fund market, we analyze the performance of US-marketed equity funds with a global geographic scope according to the Refinitiv Eikon database classification. Our analytical framework is a five-year period (2016–2021) and we combine three different databases to obtain comprehensive qualitative and quantitative data, while improving the accuracy and consistency of the data.

By using the Extreme Gradient Boosting algorithm, a robust machine learning technique, we build a reliable prediction model capable of identifying the most relevant fund features that anticipate fund performance. In this model, Environmental, Social and Governance (ESG) rating is established as a relevant determinant of fund performance. In addition, the positive association found between ESG rating and mutual fund performance definitely challenges the common preconception of a prevailing trade-off between sustainability and financial performance, and it has important implications for both investors and financial advisors.

The remainder of this paper is organized as follows. In the “Literature Review” section, the existing literature on mutual fund selection and sustainability is reviewed and the research goal is stated. The “Data Description” section then describes the sources of data and the characteristics of the sample. In the “Methodology” section, we discuss the data characteristics, variables, and model specification. The empirical results are then explained in the “Results” section. Finally, the main findings, implications, and limitations of our study are put forward in the “Conclusions”.

2 Literature Review

Researchers and academics have explored the relationship between ESG and financial performance for decades and most studies found positive correlations between ESG scores and operational efficiencies, stock performance, and lower cost of capital. High ESG ratings are now considered leading indicators of the kind of corporate transparency and high-quality management that pave the way for higher returns. Recent studies suggest that ESG investing can, under certain conditions, help improve risk management and lead to returns that are not inferior to returns from traditional investment strategies.

A number of studies have found that ESG investing may provide effective protection during economic downturns. In this sense, a recent study on the role of ESG performance during the worldwide financial crisis caused by the COVID-19 global pandemic found that companies with higher ESG scores outperformed those with lower ESG scores (Broadstock et al., 2021).

The use of machine learning methods in the finance industry has gained significant momentum in recent years. Li and Rossi (2021) highlight that fund performance exhibits a non-linear relationship with fund characteristics, emphasizing substantial interactions among various fund attributes. Their predictive modeling underscores the significance of these non-linearities and interactions. Employing advanced machine learning techniques, particularly Boosted Regression Trees, proves highly effective, surpassing the performance of standard linear frameworks. Notably, the forecasts generated by Boosted Regression Trees encompass and outperform existing predictors of mutual fund performance proposed in the current literature. Kanade et al. (2022) emphasized the use of machine learning algorithms for predicting the most suitable mutual fund scheme for investors. Their research specifically concentrated on chosen Indian companies and their respective mutual fund schemes. The experiments conducted revealed that the Artificial Neural Network demonstrated the highest accuracy in prediction.

Kaniel et al. (2023) employ neural networks for predicting mutual fund alpha, incorporating a broad range of predictors such as stock characteristics, fund characteristics, and macroeconomic variables. Their study reveals that the inclusion of fund characteristics as predictors renders stock characteristics ineffective in predicting alpha. The study conducted by DeMiguel et al. (2023) establishes that employing machine learning techniques enables the creation of portfolios for mutual funds. These portfolios exhibit positive and substantial out-of-sample risk-adjusted returns after accounting for fees. DeMiguel et al. (2023) focus on predicting abnormal returns, exclusively relying on fund-specific characteristics without incorporating macroeconomic information. They highlight the necessity of including macroeconomic variables as predictors in their model to capture the dynamics of the model under varying macroeconomic conditions. This approach aims to enhance the model’s ability to adapt and provide more robust predictions by considering a comprehensive set of predictors, including both fund-specific characteristics and macroeconomic variables.

Environmental, Social, and Governance (ESG) investing refers to an investment strategy that considers environmental, social, and governance factors in investment decision-making. According to a Deloitte report, globally, the percentage of investors who apply ESG principles to at least a quarter of their portfolios jumped from 48% in 2017 to 75% in 2019 (Collins, 2020). Some papers have analyzed the relationship between the performance of investment funds and ESG rating. Abdelsalam et al. (2020) evaluated the relationship between environmental and social governance factors and sustainable investment drivers. An examination of over 1,000 studies showed a positive correlation between ESG rating and financial performance (Whelan et al., 2021). About 53% of papers focused on performance-based ESG ratings found a positive association with fund performance. However, only 26% of papers that focused on a general statement of commitment to ESG values found a positive correlation with fund financial performance. These studies suggest that there is a relationship between mutual fund performance and ESG factors, but the exact nature of that relationship may depend on the specific ESG measures used.

Ruf et al. (2019) compare the risk-adjusted returns of socially responsible mutual funds (SMRF) with funds rated by Morningstar Portfolio ESG ScoreTM grouped into low, medium, and high ratings, during 2005 to 2016 and domiciled in the United States. They find that during the period of pre-recession boom, the funds with high-ESG ratings had a significantly higher risk-adjusted performance than did funds with a low-ESG or mid-ESG rating. This trend continued during the early parts of the Great Recession. During periods of economic recovery and growth, SRMF rated low on ESG performed significantly better than highly-rated SRMFs. For medium-rated SRMFs, other than prior to the Great Recession, fund performance was not consistently significantly different from highly-rated SRMFs.

Dolvin et al. (2019) find that funds with high sustainability scores have about the same risk-adjusted returns (i.e., alphas) as other funds. Thus, Socially Responsible Investing (SRI) investors can apparently follow a social mandate without sacrificing financial performance, but also without garnering any incremental financial benefit. They find, however, that most high-sustainability funds are concentrated in the large-cap space, which implies that strict adherence to social criteria could inadvertently result in less diversified investor portfolios. They also find that funds with high Morningstar sustainability scores generally mimic those of self-proclaimed SRI funds, suggesting that the new metric opens a larger pool of potential funds for investors focused on SRI. Steen et al. (2020) analyze the relationship between Morningstar’s ESG ratings and the performance of 146 mutual funds domiciled in Norway. Dividing the sample into ESG quintiles, they find no evidence of rating level effects, nor do they find any abnormal risk-adjusted returns (alphas). However, analyzing the European categorized funds separately, they find significantly higher returns and positive alphas for the top ESG quintiles. Abate et al. (2021), using data envelopment analysis, provide evidence of the superior efficiency of funds investing in high ESG-rated securities. Indeed, investment policies based on ESG criteria seem to enjoy a competitive advantage because of their inclusion of non-financial data. Moreover, they find that funds with high ESG ratings are subject to lower ongoing charges despite their more complex screening activities and associated operating costs. Raghunandan and Rajgopal (2022) verify whether ESG mutual funds invest in firms that have stakeholder-friendly track records, and they find that socially responsible funds do not appear to follow through on proclamations of concerns for stakeholders. Xidonas and Essner (2024) introduce a portfolio optimization model based on multi-objective minimax principles, aiming to maximize risk performance across the three standard ESG investment objectives. Their research reveals that the ESG portfolios optimized through this model consistently outperform their respective market benchmarks, yielding higher risk-adjusted returns.

Finally, Doukas et al. (2022) explore the advantages of employing both traditional classification methods, including ordinal logit, ordinal probit, and linear discriminant analysis (LDA), and machine learning techniques like k-Nearest Neighbors and Support Vector Machines. Their investigation focuses on developing models to predict the performance of energy efficiency investments. The findings reveal that machine learning methods exhibit a moderate improvement in predictive accuracy compared to traditional methods.

Overall, the literature suggests that ESG integration can lead to positive financial performance under certain conditions, but the relationship between ESG and financial performance is complex and it depends on various factors. This study endeavors to contribute to the existing body of knowledge by providing a nuanced and context-sensitive examination of the relationship between ESG and financial performance. By delving into the intricacies of this relationship, this research aims to offer a deeper understanding of how ESG considerations can be harnessed to enhance mutual fund returns. Such insights are valuable for investors, asset managers, and policymakers as they navigate the evolving landscape of responsible investing and strive to integrate ethical and financial imperatives for a more sustainable and equitable future.Footnote 1

The goal of our research is to provide support to the socially responsible investor in the fund picking process. To that end, we perform an empirical analysis with machine learning techniques and Extreme Gradient Boosting is the algorithm we use to determine those fund features that best anticipate fund performance.

3 Data Description

The data was collected in 2022 from different sources. First, from Refinitiv Eikon Database (https://www.refinitiv.com) we selected US-registered Mutual equity funds with a global geographic scope (including US stocks), an asset type involving active management in shares of any geographic scope, in dollars, with uninterrupted five-year performance records, and a minimum investment of up to €10,000, in order to focus on retail funds and exclude institutional funds. To ensure the consistency of performance ratios, the sample comprises only capitalization funds; therefore, distribution funds are excluded.

Subsequently, the Morningstar database (https://morningstar.com) was used to locate each fund in the sample, linking it to a series of relevant variables that the Morningstar company associates with investment funds. First, in order to show recommendation of these funds, they are are assigned starsFootnote 2. Second, the number of years that the fund manager has been managing each fund is considered. Finally, a variable is included that quantifies the investment style of each fund.

Finally, the Citywire database (https://citywireselector.com) is used to specify whether the fund manager is mentioned in this database and to include the fund manager rating when available.

Consequently, there is a sample of 262 funds of which the following variables are available: ESGSCORE, TER, FUNDTNA, MSSTARS, STYLEMATRIX, YEARSMANAGER, CITYWIRE, CWRATING and a series of variables for a time horizon of five years: ANNUAL RETURN, ANNUAL STANDARD DEVIATION, ALPHA, BETA (Annex II).

4 Methodology

Using financial variables, we intend to build a good regression prediction model capable of identifying the most relevant features or key factors that anticipate fund performance. The objective of this regression model is to determine which features are involved in a fund that earns strong returns, while preventing overfitting and making generalizable predictions. We use the algorithm Extreme Gradient Boosting (XGBoost) to achieve this purpose, which belongs to the family of gradient boosting algorithms. This technique is particularly useful in the context of regression situations and yields high predictive models. Boosting models combine large numbers of relatively simple tree models adaptively, to optimize predictive performance; it is particularly useful and frequently used in the context of regression problems, it can improve predictions for many regression methods, and it reduces the variance of a statistical learning method (James et al., 2017). Boosting is a sequential technique that works on the principle of an ensemble, averaging many tree models, which makes it possible to achieve high accurate predictions.

Chen and Guestrin (2016) indicate XGBoost is an effective and scalable implementation of gradient boosting machine framework by Friedman (2001) and Friedman et al. (2000). Boosting consists of creating multiple trees which are then brought together to produce a single combined prediction which substantially improves regression predictions. Hastie et al. (2009) point out the progress of the trees is sequential, that is, each tree grows using information from previously grown trees attempting to correct the mistakes of the previous ones, and each tree is fitted on a modified version of the original data set; therefore, boosting the construction of each tree depends strongly on the trees that have already been grown. Accordingly, this algorithm, in each iteration, fits a new tree using the current residuals, rather than the response variable; then the algorithm adds this new decision tree into the fitted function to update the residuals, sequentially improving the model (James et al., 2017). XGBoost adds new trees until no further improvement can be accomplished, where each subsequent model attempts to correct the errors of the previous model, and these succeeding models are dependent on the previous model.

The general idea behind gradient boosting is to generate and combine multiple weak learners—which are essentially decision trees—to produce a strong rule; therefore, gradient boosting combines multiple decision trees to lower the model error. After multiple iterations, the weak learners are combined to form a strong learner that will predict a more accurate outcome. XGBoost is a tree ensemble model which contains a set of classification or regression trees. Accumulating multiple tree predictions, the tree ensemble model overcomes the limitations or flaws of single trees regarding low predictive results. Climent et al. (2019) and Carmona et al. (2019) applied XGBoost to predict bank failure in the Eurozone and the US using classification trees, however this paper applies XGBoost based on regression trees.

One of the key elements of XGBoost models is the use of regularization, which reduces overfitting compared to other boosting techniques and allows for model generalization with new data. Regularization or feature penalization controls variable weights or penalizes complexity, performing variable selection, and decreasing high-dimensional situations. XGBoost includes regularization as part of the learning objective, as contrasted with simple gradient boosting and random forests. Moreover, XGBoost works by sequentially adding predictors to an ensemble, each one correcting its predecessor; however, instead of tweaking the instance weights at every iteration, as AdaBoost does, this method tries to fit the new predictor to the residual errors made by the previous predictor (Paul, 2021).

Training an XGBoost algorithm requires the specification of some important hyper-parameters (Kuhn & Vaughan, 2021; and https://xgboost.readthedocs.io/en/stable/ (visited on December 2022):

  • mtry. The number of predictors (variables) or proportion that will be randomly sampled at each split when creating the tree models.

  • trees. The number of trees contained in the ensemble or number of iterations or rounds.

  • min_n. The minimum number of observations in a node that is required for the node to be split further. The larger the number, the more conservative the algorithm.

  • tree_depth. The maximum depth of the tree or number of splits. Increasing this value will make the model more complex and more likely to overfit.

  • learn_rate. The learning rate, or shrinkage parameter, at which the boosting algorithm adapts from iteration to iteration. It is a small positive number that determines the contribution of each tree to the growing model. Typical values are 0.01 or 0.001. To achieve a good performance, a very small value of this parameter can require a very large number of trees.

  • loss_reduction. The minimum reduction in the loss function required to split or make a further partition on a leaf node of the tree. The larger it is, the more conservative the algorithm and the higher the regularization.

  • sample_size. The number of observations or proportion exposed to the fitting routine. Subsampling occurs once in every boosting iteration.

It is not possible to estimate the best values of these hyperparameters from data; they have to be estimated by the practitioner according to his experience or through tune techniques, such as cross-validation. Hyper-parameter tuning consists of detecting the set of the best hyper-parameters that produce a model with the highest performance on unseen data. To discover the best composition, it is quite usual to create a Grid, that is, a list or range of the most suitable values for each hyper-parameter, and then train all possible models resulting from all parameter combinations. Sometimes, when there are too many parameters’ configurations, the training process is very time-consuming and the practitioners abort the process after reaching an optimal result.

loss_reduction (or gamma) is the regularization parameter that characterizes the XGBoost algorithm from other gradient boosting techniques—XGBoost is a regularized version of gradient boosting which prevents overfitting by shrinking the leaves weights to make the boosting process more conservative. On the other hand, the decision on the number of trees or iterations (trees) is very important; in our study, we will take a value of 1,000, because it is very unlikely to get an improvement after this number of iterations as adding more trees beyond a limit does not enhance the model performance.

As mentioned above, we employ cross-validation methods to determine an optimal combination of the XGBoost hyperparameters; in particular, we choose 10-fold cross-validation (k = 10). Training data is randomly partitioned into ten subsets or folds of equal size, obtaining an overall estimation of out-of-sample error for each parameter configuration. Using the best parameter selection yielding the highest performance, a new and final model containing all training data is then fitted.

As usual in these type of Machine Learning studies, we divide the available data of investment funds into two different sets: 80% of the observations to train the XGBoost model to find the best model’s parameters configuration and the remaining 20%, as a hold-out sample, to the test data or to measure the performance of the best fitted model on independent data not used to train the model.

All models were fitted in R (R Core Team, 2022) version 4.1.3, and for XGBoost using h2o package version 3.36.0.1 (LeDell et al., 2022) and tidyverse package version 1.3.2 (Wickham, 2022). Additionally, for model interpretability, we used DALEX package version 2.4.0 (Biecek et al., 2022).

5 Results

As a precautionary measure, we ensure there is no correlational problem among the different predictors and the response variable. The size of the dots and the absence of an intense color indicates the nonexistence of this potential problem (see correlogram plot in Fig. 1). A correlogram is a graph that illustrates the correlation matrix, representing each correlation coefficient with a dot color and size according to its value.

Annex III offers a brief summary of the fundamental descriptive statistics, providing a condensed overview that captures essential features of the dataset. These statistical measures offer valuable insights into the dataset’s characteristics, aiding in a more complete comprehension of the variables and their variations.

Fig. 1
figure 1

Correlogram plot of all predictors

Full size image

We tune and fit a model considering all variables described in Annex II. The objective is to distinguish the most relevant variables, with the highest predictive content in investment funds. This process requires the ideal values of the hyper-parameters mentioned above to be identified to produce a good XGBoost model. To measure model performance, we take different measures related to model residuals, which are evaluated on an independent test set or hold-out sample, such as root mean squared error or coefficient of determination (R2). Therefore, fitting a model—while avoiding over-fitting—that could be generalized on independent data not used during the training process is crucial to tune and choose the best XGBoost model parameters’ values.

Using the R h2o package (LeDell et al., 2022) and using 10-fold cross-validation (k = 10), we explore the optimal model’s hyper-parameter configuration on the training data (80% of all data). Table 2 shows the parameters’ values for the best model after an iteration search process of 100 combinations.

Table 2 XGBoost cross-validation hyper-parameter tuning, k = 10
Full size table

The best XGBoost model—fitted taking the hyper-parameters from Table 1—returns the following resample results using the aforementioned 10-cross-validation: a quite reasonable R2 of 0.618 and a low root mean squared error of 2.039, which prove that the model is quite good. Relating to the hold-out sample or testing data (remaining 20% of all data), we even obtained a high R2 of 0.694 and also a low root mean squared error of 2.150. That the latter results are not really worse than the former clearly indicates that over-fitting is not present on the final fitted XGBoost model.

Additionally, Fig. 2, left panel, displays predicted and actual funds’ values for the test data; as most of the dots are close to the diagonal line, model residuals are quite low, confirming the goodness of fit. The right panel of Fig. 2 shows a smooth regression curve fitted for all residuals; as the result is a relatively straight line, it implies a very good model performance.

In fact, most of the residuals have very small values, as can be observed in the reverse cumulative distribution of residuals’ absolute value (Fig. 3, left panel) and in the boxplot of residuals’ absolute value (Fig. 3, right panel).

Fig. 2
figure 2

Predicted vs. actual values on test data. XGBoost model

Full size image
Fig. 3
figure 3

Reverse cumulative distribution of residuals’ absolute value (left panel) and boxplot of residuals’ absolute value (right panel). XGBoost model on test data. Note: Red dot represents the mean value of the absolute value of the residuals

Full size image

5.1 Model Interpretability

To help to understand the results of the model, that is the relationship between the predictors and the prediction target, the global interpretation of the model usually provides interesting insights. In this way, Fig. 4 shows the features that have a higher contribution to the XGBoost fitted model for fund financial performance. The importance of these features is calculated through Breiman’s permutation method (2001). The values of an important feature in the training data are randomly permuted, producing a degradation of the training performance, and destroying any relationship between that feature and the response variable, which facilitates the identification of the most important features.

Fig. 4
figure 4

Variable importance. XGBoost model

Full size image

As can be noted, our XGBoost model has revealed that the variable ESG is one of the most important variables impacting mutual funds’ performance. Our findings suggest that ESG is one of the key factors to be prioritized to improve the financial performance and long-term sustainability of mutual funds.

Besides the feature importance, another popular tool for a model’s global interpretability is the partial dependence plot. This is a model visualization technique that provides graphical insights into the prediction behavior of a machine learning algorithm, such as XGBoost. These kinds of graphs show how the expected model response behaves as a function of a selected feature, which is an average profile for all observations (Biecek & Burzycowski, 2020). Although partial dependence plots are not a perfect representation of the captured effect and while this may not provide a comprehensive description, they can show overall trends and provide a valuable basis for interpretation, explaining the marginal effect of each one of the predictors on the predicted variable, after accounting for the average joint effect of the other predictors (Friedman, 2002). Accordingly, and following Hall and Gill (2019), partial dependence plots offer a global interpretation of how a model’s predictions vary based on certain features, showing the average manner in which machine-learned response functions change based on the values of one or two input variables of interest, while averaging out the effects of all other input variables. These types of illustrations provide an easy and intuitive explanation that has been noted as practical for different purposes (Natekin & Knoll, 2013).

Figure 5 shows the partial dependence plots for the five most important variables in the final XGBoost model. The partial dependence plots indicate a positive relationship between the dependent variable (fund performance) and the following independent variables: Alpha, Morningstar fund rating (number of stars), Annual Standard Deviation (as long as volatility is under 20%), fund size (measured in total net assets), and ESG score. In addition, the plots suggest a negative relationship between fund performance and Annual Standard Deviation when volatility exceeds 20%. Consequently, the results are rather consistent with the economic interpretation of the variables.

Within the framework of our study, the prominence of the ESG factor emerges as a fundamental determinant of mutual fund performance. The empirical evidence derived from our model supports the notion that mutual funds boasting robust ESG ratings are predisposed to exhibit superior financial performance. This substantiates our hypothesis, aligning with the anticipated positive relationship between elevated ESG scores and the financial prowess of mutual funds.

The association between ESG ratings and financial performance underscores the multifaceted impact of environmental, social, and governance considerations on investment outcomes. By promoting strong ESG practices, mutual funds not only exhibit a commitment to sustainable and responsible investing but also position themselves favorably within the competitive financial landscape.

Fig. 5
figure 5

Partial dependence plots

Full size image

In addition to global interpretability, there is also a need to understand a model or predictions for a single row of data or a group of similar rows, which is known as a model’s local interpretability. Instance-level or local approaches help understand how a model produces a prediction for an individual observation. Local explanations enhance understanding by creating accurate explanations for each observation in a dataset. The principal idea is to estimate the contribution of an explanatory variable to the model’s prediction as a shift in the expected model response after conditioning on other variables. To this end, break-down plots provide a detailed summary of the effects of each particular feature on the expected model response that is easy to understand (Biecek & Burzycowski, 2020). For a particular observation of the considered funds, the XGBoost predicted value is broken down into the impact of each individual feature.

Note that the MSSTARS partial dependence plot includes six types of funds, unrated funds which are assigned zero stars and funds with 1 to 5 stars. The fact that we assigned zero stars to unrated funds explains that their performance can be higher than that of funds rated with 1 to 3 stars.

In Annex IV, we show the use of break-down plots, regarding two observations from the test set, in particular, a prediction with a high value of the response variable, and a prediction with a low value of the response variable.

6 Conclusions

In pursuit of the objectives outlined in this study, we carried out a comprehensive empirical analysis utilizing the robust machine learning technique known as Extreme Gradient Boosting (XGBoost). The primary aim was to unravel the multifaceted relationship between mutual fund characteristics and performance, with a particular focus on the essential role played by Environmental, Social, and Governance (ESG) ratings as a key determinant of fund performance.

Our research has yielded significant findings that underscore the profound influence of ESG ratings within the complex landscape of mutual fund performance prediction. ESG score emerged as one of the top five most influential factors in predicting mutual fund performance, thus affirming its status as a crucial dimension of the modern investment landscape. This finding highlights the potential for investors to align their financial aspirations with their commitment to sustainability and ethical investing without any compromise on financial returns.

The implications of our findings extend far beyond the boundaries of this study. For socially responsible investors, this revelation offers a clear path to harmonizing their values with their investment choices. It empowers them to make investment decisions that not only support their ethical principles but also align with their financial objectives. Additionally, asset managers, financial advisors, and market participants should take heed of this discovery, recognizing the increasingly prominent role of ESG considerations in shaping the investment environment. The integration of ESG criteria into investment strategies is no longer merely a matter of ethical alignment but an important factor for optimizing financial performance.

Therefore, our research underscores the significance of ESG ratings as one of the key factors in predicting mutual fund performance. This discovery dispels the notion of a trade-off between sustainability and financial returns, signaling a transformative shift in the investment paradigm. As we navigate an era where ethical and financial imperatives converge, this study serves as a catalyst for responsible investing practices, driving us toward a future where sustainable investment choices are synonymous with financial prudence, ultimately advancing both the goals of investors and the broader global community towards a more equitable and sustainable future.

In summary, this finding carries substantial implications for investors, asset managers, and financial advisors seeking to optimize their investment strategies. It emphasizes the potential for ESG integration to serve as a facilitator for financial success, offering a compelling case for the alignment of ethical and financial objectives. Moreover, our research underscores the need for further exploration into the complex mechanisms through which ESG factors exert their influence on mutual fund performance, thus providing a fertile ground for future investigations in the realm of sustainable finance.

A remarkable finding is that Alpha is the variable that best predicts fund performance by showing the impact of manager skill on fund performance. On the contrary, no relationship is found between fund total expenses and fund return. Therefore, better fund management does not necessarily involve higher management costs.

An interesting outcome is the positive relationship observed between agency ratings and fund performance. Morningstar’s rating system seems to have a relevant explanatory capacity for fund performance. Conversely, the fact that a fund manager is rated or even listed in the Citywire database does not have a significant explanatory power for fund performance.

Fund volatility, as measured by the annual standard deviation of fund returns, explains fund performance and investors willing to take some fund volatility will benefit from higher fund returns. However, as return volatility exceeds 20%, fund return declines.

All things considered, the findings of this research will help both investors and advisors focus on those fund characteristics that best anticipate fund performance. In addition, conscientious investors with a strong commitment to sustainability will find it easier to pick performing funds.

The main limitations of this study relate to the type of data and the nature of the sample. First, the empirical analysis is based on cross-sectional data and as longer data series of ESG ratings become available more robust studies may be done with time-series data for periods of market volatility, downturn, and expansion. Second, the funds comprising our sample are US-registered equity funds with a global geographic scope and further research on other fund categories is required to confirm the generalization of our results. Moreover, by selecting funds with uninterrupted five-year performance records there may be some survival bias in the mutual fund sample. An alternative approach to handle funds with missing data might improve the reliability and robustness of results. Finally, the number of sustainable funds in Europe significantly exceeds the number of sustainable funds in the US. For this reason, we look forward to exploring the performance of sustainable mutual funds in Europe in future research works.