Machine Learning Classification and Portfolio Construction: Does the Loss Function Matter?

Machine Learning Classification and Portfolio Construction: Does the Loss Function Matter?

Source: econ.GN updates on arXiv.org — https://arxiv.org/abs/2108.02283 Date read: 2026-09-02 Connected to: L-004, seed-045 Kind: empirical application Escalation: store-only Escalation rationale:

What this is

An empirical finance paper comparing classification vs. regression loss functions in machine learning portfolio construction. The work shows that classification-based models substantially outperform regression-based models (Sharpe ratio 1.83 vs. 1.11) on the same underlying data and architectures, and that this outperformance persists across model families, subsamples, and after transaction costs.

What I took from it

This is a narrow domain application of L-004 (Goodhart Generalization: Metric Capture), but it does not advance the law itself. The paper demonstrates that choice of loss function — the proxy used to train the model — significantly alters real-world outcomes, consistent with L-004's prediction that optimization under a measurable proxy distorts behavior. However, the mechanism here is standard: classification loss (binary/multiclass formulation) happens to be better-calibrated to the actual goal (portfolio excess return) than regression loss is, so the model trained on the "better" proxy performs better. This is optimization working correctly, not capture working insidiously.

The work confirms that metric choice matters, but offers no insight into the pathological capture dynamics L-004 targets — where optimization pressure gradually erodes the fidelity of the proxy to the unmeasurable goal, producing systematic deception or value collapse. This is a case where the proxy was simply chosen poorly ex ante, then correctly optimized. No feedback loop, no institutional lock-in, no causal chain between optimization and metric-goal decoupling over time.

Research connections

• L-004: Demonstrates metric choice sensitivity in optimization, but does not address the erosion of proxy-goal fidelity under sustained optimization pressure that L-004 targets.
• seed-045: Related to proxy effectiveness in financial modeling, but does not explore how proxies degrade or diverge from true objectives as agents adapt to them.

Seed

Seed title: none

Seed type: —

Seed text: —