Research Papers
Operational Validation of Continuous Time-Series AI Systems — IEEE
ACDSA 2027 – peer reviewed
Date Written: Sept 2026
Historical replay of predictive AI can be causally invalid when archived data do not represent the information actually available at runtime. This paper reports a forensic validation case study from an operational AI-assisted foreign-exchange platform. The study distinguishes canonical candle close from operational availability, uses an independent integrity observer to monitor candle completeness and timing, and evaluates replay/runtime consistency as bounded contracts rather than a single binary property. A frozen pre-flight identified 130 missing lower-timeframe candle observations across 12 symbols; a later detector study covering 11,543 expected multi-timeframe candle observations classified 2,294 (19.87%) as late, one as missing, and 116 as timing-unproven. A preserved configuration-authority integration check reported 115 focused tests passed but authorized no routes and therefore does not establish end-to-end trading-decision parity. The results show that causal data state and replay/runtime equivalence require explicit validation before operational AI performance is interpreted.
Rapid-Iteration Multi-AI Trading Platform: Methodology and Mid-Stage Results on EURUSD
Date Written: September 09, 2025
The development of profitable, reliable trading systems remains a persistent challenge in financial markets, where dynamic conditions, noise, and unexpected events can quickly
invalidate static models. This paper presents the design and implementation of a multi-stage, multi-AI trading platform that integrates feature engineering, ensemble modeling, and
automated validation to adapt to changing market conditions in real time. Unlike single-model approaches, the platform employs multiple AI engines working in sequence to generate, check,
and refine trade signals before execution. Initial research focused on EURUSD, chosen for its volatility and liquidity during development, which created both setbacks and opportunities.
Failures uncovered by rapid iterations were not treated as dead ends but as critical feedback, enabling accelerated improvement cycles. Built on rapid prototyping principles—multi-hour
iterations rather than multi-week sprints—the system demonstrates how quick identification of weaknesses can ultimately drive robustness. Results to date indicate that the platform can
identify high-probability entry/exit conditions, reduce failure rates through built-in self-checks, and provide a scalable framework for multiple symbols and strategies. Future work includes
reinforcement learning integration, cross-timeframe attention models, and automated risk calibration to further improve resilience and profitability.