Noise and Sequence Length in Time-Series Forecasting: A Unified Evaluation of Transformers, DLinear, LSTM, and GRU

Intelligent environments in critical domains require forecasts that remain reliable under noisy sensing and changing operating regimes. This work presents a two-factor evaluation that varies the input window length and training noise level across multiple forecast horizons, four model families (iTransformer, LSTM, GRU, and DLinear), and 13 datasets spanning lowand high-frequency signals. A unified preprocessing pipeline with multi-seed reporting quantifies accuracy (relative MSE and relative MAE) and stability. Three main findings emerge: (i) Forecast horizon dominates, as shorter horizons consistently reduce relative error; (ii) Long noisy windows degrade most, with error and uncertainty increasing more than additively when long input windows are combined with non-zero noise; and (iii) Architecture matters under noise, as attention-based and gated models mitigate degradation, whereas a single linear head is most vulnerable. The study provides a reproducible benchmark across timescales and practical guidance: favor short horizons, use long input windows only under clean sensing conditions, and stresstest forecasting pipelines with realistic noise.