TRN: A 36K-Parameter Oscillator Network for Time Series Forecasting

Introduction: Transformer-based architectures dominate time series forecasting but incur substantial parameter and computational costs, limiting deployment in resource-constrained and edge settings. We present TRN (Temporal Resonance Network), a compact neural architecture that replaces attention with a learnable bank of 48 damped oscillators.

Methods: TRN represents temporal dynamics through oscillators whose frequencies are geometrically initialized (0.005–0.25 Hz) and jointly learned with damping coefficients. Oscillator responses are computed via FFT-based convolution and combined through a frequency-mixing layer, followed by a shared projection with per-oscillator rank-4 LoRA adapters. An optional Student-t head provides calibrated uncertainty, and the oscillator energy representation additionally enables zero-shot anomaly detection.

Results: On eight standard benchmarks (ETTh1/2, ETTm1/2, Exchange, Solar, Electricity, Traffic), TRN achieves MSE comparable to much larger baselines using only ~36K parameters—an order of magnitude fewer than typical transformer architectures (1–7M parameters). At horizon P=48, TRN attains 0.375 test MSE on ETTh1, 0.064 on Exchange, and 0.164 on Solar (3-seed mean). The Student-t head yields well-calibrated prediction intervals with under 1% point-forecast degradation. For anomaly detection, the oscillator energy representation achieves 0.85 AUROC on the SMD benchmark without any detection-specific training, matching dedicated deep anomaly-detection methods that require explicit supervision.

Conclusions: TRN demonstrates that oscillator-based dynamics offer a parameter-efficient alternative to attention for time series forecasting, with built-in support for probabilistic prediction and zero-shot anomaly detection.