Machine Learning and Fund Characteristics Help to Select Mutual Funds with Positive Alpha

Published: 3 September 2026| Version 4 | DOI: 10.17632/rpgb99m5zy.4
Contributors:
Andre Santos, Victor DeMiguel, Javier Gil-Bazo, Javier Nogales

Description

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Replication files to ``Machine Learning and Fund Characteristics Help to Select Mutual Funds with Positive Alpha'' %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Important notice: Fausch et al. (2025, SSRN link: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5937794) identify an error in this replication code that introduces an unintended look-ahead bias in the construction of the mutual-fund portfolios. Specifically, in the computePortfolioReturns function, line 67 of code_for_ML_methods.Rmd, line 49 of code_for_EW_method.Rmd, line 55 of code_for_AW_method.Rmd, and line 53 of code_for_table_6.Rmd update portfolio weights using fund.returns[, r+1] when they should use fund.returns[, r]. Correcting this error eliminates the long-only portfolio's significantly positive alpha, although the corresponding long-short portfolio still delivers significant alpha; the paper's other main conclusion — that machine learning identifies managers whose skill is not fully offset by diseconomies of scale — is unaffected. DeMiguel, Gil-Bazo, Nogales, and Santos (2026, SSRN link: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6860698) show that the long-only result can be restored by refining the estimation procedure: using a quarterly rather than annual panel (more observations to train the models), Bayesian optimization to select hyperparameters, and quarterly rebalancing. See Fausch et al. (2025) and our corrigendum for details. Each folder contains files to replicate one table or figure of the paper. The folders also contain a readme file to help users to run the codes. The replication_file.tex file generates a PDF file with the replicated tables and figures. Notice that the replicated tables and figures are not always identical to those reported in the paper. The reason is that in order to protect the proprietary nature of the data, in some cases we have added noise to mutual fund characteristics and returns; see the README file for details. The data files that are required to run the codes are stored in the /data_sets/ folder. Some tables and figures require the output from the codes stored in folders /code_for_ML_methods/ , /code_for_AW_method/ and /code_for_EW_method/. It is recommended to run the codes in those folders before running the codes that generate tables and figures. We include a file in both tex and pdf formats containing the full set of results.

Files

Institutions

Categories

Investment, Financial Econometrics, Machine Learning

Funders

Licence