Hopformer cuts time series forecast error by 6.56% with two-stage approach

Forecasting multiple time series with many covariates is notoriously difficult: models must capture shared global patterns while preserving unique series-specific dynamics. The newly proposed Hopformer (Homogeneity-Pursuit Transformer) tackles this by splitting the problem into two distinct stages. First, a Sparsity Pattern Aggregation (SPA) scheme extracts a common, low-variance trend that incorporates covariates—effectively a homogenization layer that reduces noise and stabilizes the signal. Then, a LoRA-fine-tuned Transformer models the remaining complex dependencies in the residual. The authors prove that SPA achieves a near-optimal bias-variance trade-off via an oracle inequality, and derive generalization bounds for the second stage under dependent time series data.

On benchmarks, Hopformer delivers a 6.56% average improvement in MASE (Mean Absolute Scaled Error) over previous state-of-the-art methods, spanning both synthetic and real-world forecasting tasks. The framework is theoretically sound and practical, requiring only lightweight LoRA fine-tuning for the transformer component. For professionals in finance, energy, or IoT—where high-dimensional time series are the norm—Hopformer offers a significant accuracy lift without massive computational overhead. The work is available on arXiv (2607.22299) and represents a promising new direction for robust, large-scale forecasting.