New research debunks the “Virtue of Complexity in Return Prediction” in finance
In "Simplified: A Closer Look at the Virtue of Complexity", Professor Daniel Buncic of Stockholm Business School has released a new study that disputes the findings of one of the most hotly debated papers in quantitative finance this summer.
See the articles in the Financial Times and Bloomberg on this topic.
Daniel Buncic's research challenges the claim by Kelly, Malamud, and Zhou (2024, KMZ) in "The Virtue of Complexity in Return Prediction" that models with thousands of parameters trained on a short span of financial data can outperform simpler benchmark models. While KMZ’s research has attracted considerable attention in both academia and industry, Daniel Buncic highlights key methodological flaws that undermine its broader implications for systematic investing.
In particular, he shows that the supposed "Virtue of Complexity" is an artifact of two implementation choices made by KMZ which make the performance of simpler models appear much worse than they really are. First, KMZ restrict the intercept term in their prediction models to zero. Second, they employ a paradoxical aggregation scheme in the construction of the complex model’s performance measures that no investment professional would (or could) ever implement. These two design choices lead to the seemingly poor performance of simple models.
Professor Daniel Buncic. Photo: Private
Daniel’s study demonstrates further that a simple expanding-window linear regression model with mild shrinkage achieves a Sharpe ratio of 0.699, which is significantly higher than the 0.485 of KMZ’s preferred most complex model. Even historic averages, often dismissed as naïve, outperform the complex machine learning alternatives proposed by KMZ.
Complexity does not appear to be a virtue, at least not for return prediction.
Read the paper "Simplified: A Closer Look at the Virtue of Complexity"
Read the paper "The Virtue of Complexity in Return Prediction"
Last updated: 2025-12-02
Source: SBS