Quantitative Monographs "Extending Quality with ML using a composite sig..."

Research evidence excerpt

Quantitative Monographs "Extending Quality with ML using a composite sig..."

Quantitative Monographs UBS Research

Executive Summary

Extending our work on Interpreting Quality using Machine Learning

In our original quality framework (Quantitative Monographs "Interpreting Quality with

Machine Learning" Antrobus), we showed that residual return prediction using quality-

related fundamentals delivers a signal with attractive long–short performance. That

work was structured as a comparison between two modelling approaches: Elastic Net

and Gradient Boosted Machines.

Elastic Net (hereafter: ENet), which imposes linear structure and strong

regularization.

Gradient Boosted Machines (hereafter: GBM), which allow for non linear

interactions and threshold effects.

Disagreement matters

Whether the models agree or disagree on future residual return expectations is an

important factor driving realised outcomes. Where the models disagree, for example

when a stock is in the top third from one and the bottom third for the other, this leads to

low returns and volatility.

Composite model leverages both models in a single score

This note builds on our previous findings, moving beyond a horse race between models

and asking whether their complementary strengths can be combined. This naturally

motivates a composite approach: ENet captures stable, linear quality premia while GBM

captures conditional and non linear effects, combining them preserves alpha while

improving robustness.

This note evaluates an equally weighted composite quality score, showing that it

dominates both ENet and GBM models as well as traditional Quality benchmarks, on

average generating 1.5% annual residual return, and 1.6% annual raw return.

FigureSource:UBS2: QuantitativeAverage Researchannual returns by modelAverage