UPLOTS: A Unified Pretrained Language Model for Constrained Time-series Generation

UPLOTS is a unified pre-trained language model that generates constrained time-series data across multiple domains using a single transformer backbone guided by learned prompts. The framework addresses scalability limitations of existing domain-specific approaches by internalizing diverse temporal structures and enabling conditional generation with precise pattern control.

UPLOTS represents a significant advancement in time-series generation by consolidating what has traditionally required separate, handcrafted models for each dataset into a single unified framework. The research tackles a fundamental inefficiency in the field: the inability to leverage shared temporal patterns across different domains. By using a pretrained transformer backbone with constraint-guided prompts, the system achieves both flexibility and scalability that previous approaches lacked. The dynamic multi-dataset loss re-weighting mechanism enables the model to learn from diverse temporal structures simultaneously, improving generalization capabilities. The framework demonstrates practical value across multiple real-world applications including peak-period forecasting, calendar-based patterns, load-level predictions, and volatility modeling. Performance validation extends beyond simple benchmarking, with held-out constraint combinations and downstream forecasting experiments confirming the model's ability to handle unseen pattern combinations and enhance data augmentation in scarce-data scenarios. For the machine learning and financial forecasting communities, this work addresses a critical pain point in deploying time-series solutions across heterogeneous environments. The ability to maintain a single production model rather than managing dozens of domain-specific variants reduces operational complexity and maintenance burden. The research particularly benefits organizations dealing with limited historical data, where synthetic augmentation through UPLOTS could improve model training quality. The public code availability accelerates adoption potential, though real-world deployment success will depend on how well the framework adapts to proprietary data distributions and emerging temporal patterns.

• →UPLOTS uses a single pretrained transformer model with constraint prompts instead of building separate models for each dataset, improving scalability and generalization
• →Dynamic multi-dataset loss re-weighting allows the framework to internalize diverse temporal structures from multiple domains simultaneously
• →The model demonstrates strong performance on peak-period, calendar, load-level, and volatility pattern generation across real-world benchmarks
• →UPLOTS improves data augmentation effectiveness in scarce-data regimes, making it valuable for organizations with limited historical time-series data
• →Held-out experiments confirm the framework generalizes to unseen constraint combinations beyond original training patterns