TIP-Search: Time-Predictable Inference Scheduling for Market Prediction under Uncertain Load

TIP-Search presents a systems-level scheduling framework for real-time market prediction that balances prediction accuracy with deadline satisfaction under computational constraints. Using constrained online optimization and a shielded expert selector (OCO-ACPO), the approach achieves 99.1% timely accuracy and 96.2% deadline satisfaction on financial order book prediction tasks, demonstrating that temporal guarantees matter as much as prediction quality in production trading systems.

TIP-Search addresses a critical gap in machine learning deployment for financial markets: most prediction systems optimize for accuracy alone, ignoring the hard constraint that late predictions are worthless. This research treats market prediction as a scheduling problem where multiple candidate models compete for finite computational resources, and the system must choose which model to run on each prediction request to maximize both correctness and timeliness. The framework introduces OCO-ACPO, a projected-dual expert selector that dynamically routes inference tasks while maintaining a safety shield against deadline violations and queue overload. On the FI-2010 order book dataset, TIP-Search improved deadline satisfaction from 39.1% to 96.2% while raising timely accuracy from 15.6% to 23.9%, a substantial engineering gain. The nonstationary variant SA-OCO-ACPO further improves performance under stress conditions, handling variable computational load without degrading service guarantees. This systems-oriented approach differs fundamentally from classifier leaderboards that assume unlimited inference time. For practitioners building trading infrastructure, the research validates that temporal scheduling policies significantly impact real-world prediction service quality. The methodology applies beyond financial markets to any time-constrained inference deployment, including autonomous systems and real-time monitoring. The statistical significance (p-values below 0.02) indicates these improvements are robust across test conditions, not artifacts of specific scenarios.

• →Market prediction systems require deadline-aware scheduling, not just high accuracy, as late predictions provide no trading value
• →OCO-ACPO achieves 96.2% deadline satisfaction and 30.3% timely accuracy by dynamically routing inference across heterogeneous models under resource constraints
• →Nonstationary stress scenarios improve 18.8-41.7 percentage points with SA-OCO-ACPO over baseline constrained optimization
• →The approach demonstrates statistically significant improvements (+0.0146 deadline satisfaction, p=1.5×10⁻⁵) over existing scheduling baselines
• →This systems result emphasizes that production trading infrastructure must co-optimize temporal guarantees alongside prediction quality