← Research overview

Improving Bayesian Optimization for Portfolio Management with an Adaptive Scheduling

Zinuo You, John Cartlidge, Karen Elliott, Menghan Ge, Daniel Gold
ICAAI 2025

Adjust variance pressure as an optimization budget is spent on an expensive, noisy black-box objective.

Lead researcher and implementation author · Jun 2024–Apr 2025

Search under a limited budget

An expensive evaluator may produce an attractive score by chance. Optimization must balance high observed utility against uncertain outcomes while spending a finite number of evaluations.

An adaptive objective

TPE-AS augments Parzen-density search with a budget-dependent mean–variance objective and clipped importance correction. The public implementation provides reusable optimizer mechanics and synthetic benchmarks.

Reading the demonstration

The toy objective has a broad, stable peak and a narrow, noisy peak. A fixed, seeded set of evaluated candidates is ranked by sample mean minus a scheduled variance penalty. Move the budget-progress control to see the selected incumbent change. This isolates adaptive scoring; it does not implement Parzen proposal generation, importance correction, or a financial backtest.

Interactive toy experiment

A high score can also be a noisy score

Move through the budget schedule to rerank the same candidates. Change noise or seed to inspect different observations.

A high score can also be a noisy score-0.20.01.50.53.21.0Candidate parameter x

Gray line: true mean · Blue dots: sample mean · Dashed: penalized score · Ring: selected

Variance penalty λ = 0.500 · Selected x = 0.82 · Mean 2.142, variance 0.119, score 2.083. 25 candidates; 5 observations each.

How this toy works

25 fixed candidates receive 5 Gaussian observations each. Score = sample mean − λ × sample variance, with λ = [1 − cos(π × budget fraction)]/2. Select the largest current score.

This isolates adaptive scoring. It omits Parzen proposals and importance correction, and is not the complete optimizer or a portfolio backtest.

Synthetic illustration · Not a trained-model output or a published benchmark.