Trustworthy AI: Deciding What to Decide

Caesar Wu, Yuan-Fang Li, Jian Li, Jingjing Xu, Pascal BouvryarXiv 2023 · DOI · arXiv

Abstract

In strategic financial decision-making, the opacity of artificial intelligence often hinders its adoption for high-stakes tasks. This paper introduces a structural framework for Trustworthy AI (TAI) that decomposes "trust" into twelve measurable properties mapped across the machine learning pipeline: representation space, loss functions, and optimizers. Applying this framework to Credit Default Swaps (CDS) investment, the study demonstrates that traditional tree-based models like XGBoost currently outperform modern Transformer architectures in both predictive accuracy and the "trustable" explainability required for financial strategy.

Introduction: The "Deciding What to Decide" Problem

Strategic decision-makers, particularly in finance, frequently face a deluge of contradictory information. While AI offers the promise of processing this complexity, its "black box" nature often makes it difficult to determine which outputs are reliable enough for action. This is the "deciding what to decide" problem. The authors address this by formalizing trust not as a vague sentiment, but as a rigorous engineering requirement. This is especially pertinent in the context of Credit Default Swaps (CDS), financial instruments that played a central role in the 2008 financial crisis and require high levels of transparency.

The TAI Framework

The core of the research is the TAI Framework, which organizes twelve ethical and technical properties into three dimensional components of the AI lifecycle:

The TAI Framework mapping dimensional components to ethical and technical properties.
  1. Representation Space (The Model, M): Focuses on "Value" properties such as Justice, Explainability, Transparency, and Fairness.
  2. Loss Function (The Data, D): Focuses on "Evaluation" properties including Availability, Usability, Security/Privacy, and Accountability.
  3. Optimizer (The Algorithm, A): Focuses on "Selection" properties like Robustness, Reproducibility, Reliability, and Accuracy.

Experimental Methodology

To validate the framework, the authors analyzed a massive CDS dataset (2006–2017) comprising nearly 750,000 observations across 10 industrial sectors. The study specifically focused on the Technology Sector, utilizing 117 features to predict CDS spreads.

The researchers compared two primary classes of models:

  • Tree-Based Models: Gradient Boosting Machine (GBM) and XGBoost (Xgbm).
  • Transformer Models: TimesNet, PatchTST, and Crossformer.

To ensure the "Value" property of explainability was met, they employed an Explainable AI (XAI) suite including SHAP, LIME, and Partial Dependence Plots (PDP).

Key Results and Insights

The experiments yielded several critical findings for the application of AI in finance:

  • Performance Superiority of Tree Models: XGBoost achieved the highest accuracy with a Root Mean Square Error (RMSE) of 26.30, significantly outperforming the best Transformer model, PatchTST (RMSE 38.29).
  • The Explainability Gap: While Transformers are state-of-the-art in many domains, they lack the mature, stable interpretability tools available for tree-based models. In finance, where understanding the "why" is as important as the "what," this gap makes Transformers less "trustworthy" for current strategic use.
  • Feature Stability: Across all models, "Equity Value" and "Price Sale" were identified as the most influential predictors. This cross-model consistency allows human decision-makers to place higher trust in these specific indicators.
Formal representation of the trust relationship where Environment (E), Data (D), Algorithm (A), and Model (M) satisfy Properties (P).

The authors argue that trust is a structural outcome of the interaction between the Environment (E), Data (D), Algorithm (A), and Model (M). When these elements satisfy the required Properties (P), the system can be deemed trustworthy.

Conclusion and Future Directions

The study concludes that current AI operates primarily as "System 1" thinking—fast pattern matching without deep logical reasoning. To achieve true trustworthiness, future AI must integrate "System 2" capabilities, combining modern machine learning with rule-based logical systems like Cyc. While tree-based models currently lead in the financial domain due to their balance of speed and transparency, the TAI framework provides a roadmap for evolving more complex architectures into reliable decision-support tools.

Summary updated

Abstract

When engaging in strategic decision-making, we are frequently confronted with overwhelming information and data. The situation can be further complicated when certain pieces of evidence contradict each other or become paradoxical. The primary challenge is how to determine which information can be trusted when we adopt Artificial Intelligence (AI) systems for decision-making. This issue is known as deciding what to decide or Trustworthy AI. However, the AI system itself is often considered an opaque black box. We propose a new approach to address this issue by introducing a novel framework of Trustworthy AI (TAI) encompassing three crucial components of AI: representation space, loss function, and optimizer. Each component is loosely coupled with four TAI properties. Altogether, the framework consists of twelve TAI properties. We aim to use this framework to conduct the TAI experiments by quantitive and qualitative research methods to satisfy TAI properties for the decision-making context. The framework allows us to formulate an optimal prediction model trained by the given dataset for applying the strategic investment decision of credit default swaps (CDS) in the technology sector. Finally, we provide our view of the future direction of TAI research
BibTeX
@misc{wu2023trustworthy,
  author = {Caesar Wu and Yuan-Fang Li and Jian Li and Jingjing Xu and Pascal Bouvry},
  title = {{Trustworthy AI: Deciding What to Decide}},
  year = {2023},
  eprint = {2311.12604},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2311.12604},
}
Figure