ML in Production · 3 min read
Fifteen ML models beat the S&P 500 — then profitability decayed
An Information Systems Frontiers editorial synthesizes studies showing fifteen ML models beat the S&P 500 with decaying profitability, and argues explainability and governance must be baseline system components.
Key takeaways
- Olson et al. tested fifteen ML models on S&P 500 prediction; profitability declined markedly over time.
- Editorial published in Information Systems Frontiers, DOI 10.1007/s10796-026-10846-9, 2026.
- Spline-rule ensembles matched predictive performance while preserving interpretable rules in credit scoring.
- Editors warn homogeneous models and shared datasets create model concentration risks during market volatility.
- Research agenda covers agentic architectures, federated analytics, and explainable access controls.

Fifteen machine learning models generated abnormal returns predicting the S&P 500 — and their profitability declined markedly over time. That finding anchors an editorial published in Information Systems Frontiers (DOI: 10.1007/s10796-026-10846-9), where Konstantina Spanaki, Christos Floros, Ioannis Chatziantoniou, Constantin Zopounidis, and Christos Lemonakis synthesize a collection of studies on AI in financial analysis and data sharing. Their core verdict: predictive accuracy does not reliably convert into stable economic value, and analytical efficiency cannot be the sole criterion for evaluating intelligent systems.
What did the empirical work actually show?
The study by Olson and colleagues evaluated fifteen machine learning models against technical analysis data for S&P 500 prediction across an extended historical period. Returns varied substantially across intervals. The decline over time is the part vendors rarely quote; it suggests model decay under evolving market conditions is the rule, not the exception. Data teams running market-facing models should budget for retraining and degradation monitoring from day one.
A second study, by Jha and coauthors, applied the Technology Acceptance Model (a survey framework predicting technology adoption) to finance professionals. Four drivers dominated: effort expectancy, performance expectancy, trust, and organizational innovativeness. In other words, adoption is an organizational variable, not a benchmark score.
Can explainability survive competitive accuracy?
De Lange and colleagues tested spline-rule ensembles (interpretable model structures built from simple conditional rules) on credit scoring. The models achieved competitive predictive performance while preserving readable rule structures. That result chips at the standard excuse that transparency must be traded against power.
An experiment by Ioannou and coauthors added a caveat. User adoption of automated recommendations depended on the interaction of explanation quality, perceived decision risk, and the nature of the advisory entity. Explainability, they conclude, is not a static parameter. It must fit specific decision contexts.
What does governance look like in practice?
- Chatterjee and colleagues used adaptive structuration theory to map safeguards for generative AI in healthcare data sharing within financial contexts.
- Lioliou and coauthors analyzed risk disclosures from publicly listed financial institutions, finding accountability is constructed through corporate narratives and formal disclosure, not only internal technical controls.
- Oikonomidis and colleagues proposed regulatory architectures for auditing dynamic residential electricity pricing algorithms while minimizing disclosure of sensitive consumption data.
The editors' position is blunt: fairness, transparency, and accountability gain practical meaning only when embedded in verifiable controls, audit procedures, and structured workflows — not appended retroactively at the architectural level.
Where are the systemic risks?
Two concerns stand out. First, model concentration: homogeneous models trained on shared financial datasets create interdependencies that could destabilize markets during acute volatility. Second, agentic ambiguity: as generative and multi-agent systems produce forecasts, recommendations, and synthesized narratives, assigning accountability for a specific outcome becomes harder when multiple autonomous models interact with human analysts.
Fagbola and colleagues illustrated behavioral fragility empirically, applying topic modeling and sentiment analysis to platform X discussions around the NFT market crash. Emotional dynamics, social identity, and herd behavior evolved through successive phases of instability. A bibliometric study by Bawack and coauthors mapped AI, accounting, and sustainable finance scholarship, flagging gaps in sustainability risk assessment, regulatory technology, and environmental governance.
What should practitioners take away?
The editorial insists AI value depends on organizational readiness, leadership, user acceptance, and institutional context rather than computational sophistication alone. Prior research on fintech adoption shows managerial capabilities frequently decide digital initiative outcomes. The recommended research agenda includes accountable systems, agentic architectures, cross-organizational data ecosystems, federated analytics (a privacy-preserving method that trains models without pooling raw data), and explainable access controls.
The skepticism is warranted. The collection spans commercial banking, credit scoring, risk governance, healthcare data collaboration, energy pricing, and supply chains, using methods from econometrics to controlled experiments. That breadth is a strength. But an editorial synthesizing invited studies cannot settle whether the proposed governance mechanisms work at scale; longitudinal evidence does not yet exist. Treat the framework as a hypothesis set, not a standard.
doi.org (Original)
More from Tom Whitfield
Show full bio
Staff writer covering regulation, policy and their impact on Analytics Digest is a professional daily at Analytics Digest.
16 articles