An Analysis and Improvement of Robo-Advisory

Abstract

This dissertation examines and seeks to improve robo-advisory (RA) through an integrated analysis of German providers, a dynamic formal model of the workflow, an operational large-language-model (LLM) prototype, and a benchmark-consistent performance study. It tackles opacity in profiling and portfolios, the absence of a unified dynamic formalism, the lack of a scalable, auditable LLM-driven RA, and limited risk-matched evidence on realised outcomes. Six questions frame the work: RQ1–RQ2 on questionnaires, portfolios and strategies; RQ3 on a mathematical framework; RQ4 on a ChatGPT-4o-based RA; and RQ5–RQ6 on risk-adjusted performance overall and during COVID-19.

Methodologically, a July-2024 census of 45 German RAs is used to construct a profiling taxonomy validated with text analytics features and clustering. Holdings-based mapping is combined with structured analysis of strategy disclosures. A pragmatic two-layer catalogue—eleven equity cores with horizon/environmental, social, and governance (ESG)/knowledge overlays—is proposed. The formal contribution treats profiling, allocation and event-driven rebalancing as a unified control problem. A ChatGPT-4o prototype, built around a 32-item questionnaire and an exchange-traded fund (ETF) universe, generates 100 allocations; strict validation and regression relate questionnaire items to equity share. The performance study assembles monthly indices, assigns portfolios to equity buckets, and compares them with investable benchmarks using standard performance ratios.

Findings show convergence on eight profiling themes but uneven depth: risk, horizons, experience, and finances are widely covered, yet psychometric risk and knowledge checks are often sparse; ESG preferences have grown. Recommended portfolios are dominated by global equity/bond ETFs, with equity share the main risk dial; menus remain narrow relative to questionnaire breadth; mean–variance thinking dominates, and artificial intelligence (AI) use is nascent. The ChatGPT-4obased prototype produces mixes in which equity exposure is chiefly explained by risk-attitude items and liquidity buffers; validation layers are necessary to prevent accounting/eligibility errors. On realised outcomes, RAs underperform matched benchmarks on average across buckets over 2017–2022 and during COVID-19, with higher beta/correlation, deeper drawdowns and slower recoveries explaining efficiency gaps; a minority outperform in specific buckets.

The thesis recommends a supervisory profiling taxonomy grounded in validated psychometrics, machine-readable reporting and explicit mapping from client risk grades to product risk labels; event-driven rebalancing with beta/correlation control; and transparent, bucket-matched performance disclosure. Providers should adopt two-layer menus or dynamic instantiation and deploy explainable, auditable LLM pipelines with guardrails and human oversight. For investors, clearer mappings from responses to risk labels and risk-adjusted, benchmarked reporting are essential. Collectively, the programme offers a practical blueprint to make RA more transparent, accountable and performant.