Self-supervised contrastive learning has recently shown promise for time-series representation learning, yet most existing methods treat sequences holistically and leave trend and seasonal components entangled, limiting their effectiveness for long-horizon multivariate forecasting. We study decomposition-aware representation learning for time-series forecasting without negative pairs. We propose the Trend-Season Contrastive Learner (TSCL), a Siamese framework that decomposes each series into trend, seasonality, and residual components, encodes trend and seasonality with dedicated encoders and a learnable Fourier layer, and optimizes a positive-pair contrastive objective over component-wise representations. Experiments on five public benchmarks (ETTh1, ETTh2, ETTm1, ETTm2, and Weather) show that TSCL consistently improves downstream forecasting across prediction horizons. Averaged over all datasets and horizons, TSCL achieves 0.489 MSE and 0.488 MAE, yielding an about 20–30% lower error than representative contrastive baselines (e.g., SimTS and CoST). Paired t-tests further confirm that the improvements are statistically significant in most settings. These results indicate that decomposition-aware contrastive learning yields robust and generalizable representations for long-horizon forecasting across diverse temporal resolutions.
Chou et al. (Thu,) studied this question.