Human Narratives for Suitability-Aware Robo-Advisory

Description

This dataset supports a study of suitability-aware robo-advisory using natural-language client disclosures. It contains 129 anonymized, human-authored investment narratives collected through LinkedIn, Facebook, WhatsApp, and email, together with structured investor-profile information. Thirty-nine cases include three-stage narratives that capture changes in personal, financial, or behavioral circumstances over time. Two independent coders evaluated every case across thirteen dimensions, including investment objective, time horizon, risk tolerance, risk capacity, liquidity needs, behavioral signals, contradictions, missing information, suitability risk, portfolio allocation, and escalation requirements. The workbook includes both coder datasets, all adjudicated disagreements, final labels, the codebook, and case-level predictions for the static questionnaire baseline. Summary results compare the static questionnaire, rule-based NLP, LLM-only, and controlled NLP architectures. Case-level predictions are available for the static baseline, while results for the other architectures and the multi-turn evaluation are reported as aggregates. The workbook is intended to support transparency, verification of the human coding process, and replication of the baseline analysis.

Steps to reproduce

Open the workbook in Microsoft Excel or another application that preserves Excel formulas. Begin with the README sheet, which explains the workbook structure and the fixed decision rules, and consult the Codebook sheet for the definitions and permitted values used in the coding process. The Responses sheet contains the 129 retained cases and their structured investor information. The Coder1 and Coder2 sheets contain the independent assessments of each case across thirteen coding dimensions. Compare these sheets to identify coder disagreements, then consult the Adjudication sheet for the final decision applied to each disagreement. The resulting adjudicated labels and advisory outcomes are presented in Final_Labels. To reproduce the static-questionnaire baseline, open M1_Predictions. The formulas import each structured risk score from Responses and apply the fixed rules described in the README and Codebook. Scores from 1–3 produce a Conservative portfolio, 4–5 a Balanced portfolio, 6–7 a Growth portfolio, and 8–10 an Aggressive Growth portfolio. Suitability risk is classified as Low for scores of 1–2, Medium for 3–5, and High for 6–10. The predicted outcomes are then compared with the adjudicated labels. The Results sheet calculates the retained sample size, coder agreement, field-level Cohen’s kappa, outcome distributions, narrative-information measures, baseline performance, and model-comparison statistics. The MultiTurn_Summary sheet reports the stage-level distributions and transition performance for the 39 multi-stage cases and 78 observed transitions. Recalculation of the workbook formulas reproduces the displayed results.

Institutions

• Ca' Foscari University of Venice

  Veneto, Venice

Categories

Licence