Compact time series encoder matches performance of models 54 times larger

A new time series encoder called OTIS demonstrates that effective feature extraction doesn't require massive models. Researchers have developed a 7.1M parameter system that rivals the performance of much larger counterparts while being deployable on power-constrained devices like wearables and industrial sensors.

The Scale Problem#

Most cutting-edge time series encoders rely on enormous models to capture patterns across diverse data sources. This reliance on sheer size makes real-world deployment impractical—many resource-constrained systems simply cannot afford the memory footprint, energy draw, or computational latency that large models demand. This creates a significant gap between research capabilities and practical usability.

How OTIS Works Differently#

OTIS achieves competitive performance through three key innovations tailored specifically to time series data:

• Domain-aware tokenisation handles the inconsistencies that arise when training on heterogeneous data from multiple sources
• Dual masking strategy learns both spatial and temporal patterns while preserving causal relationships in the data
• Structure-aware objective separates meaningful signal learning from noise capture

By adapting masked modelling pre-training to respect the unique properties of time series, the approach sidesteps the need for scale without sacrificing quality.

Performance and Efficiency#

Testing across 162 downstream tasks, OTIS matches state-of-the-art results from encoders that are 54 times larger. More practically, it delivers:

• 10× reduction in memory requirements
• 43× lower energy consumption
• 37× faster inference latency

The efficiency gains extend beyond discriminative tasks—OTIS features also enable generative applications with minimal overhead.

Access and Impact#

To encourage adoption, the researchers have released code and pre-trained weights, removing barriers to deploying sophisticated time series analysis on any system regardless of computational constraints.