Deep learning models such as Transformers (e.g., PatchTST) and linear-centric architectures (e.g., DLinear) achieve state-of-the-art performance on stationary benchmarks. However, real-world data streams exhibit concept drift and regime shifts that challenge static models, often leading to “silent failures” where peak error spikes dangerously. While meta-learners or frequent retraining can mitigate drift, they incur high computational costs unsuitable for resource-constrained edge environments. This paper presents the Sliding Window Selector (SWS), a training-free, computationally frugal dynamic ensemble that navigates the stability-plasticity dilemma by combining two frozen experts: PatchTST (steady) and DLinear (volatile). We introduce a rigorous statistical heuristic relying on an Exponential Moving Average (EMA) error tracker (α = 0.3) coupled with softmax-based weighting. Crucially, we propose a novel Distress Fallback mechanism that enforces a Value-at-Risk (VaR) style constraint when ensemble reliability degrades beyond a historical confidence interval. We evaluate our framework on two datasets (Energy and Finance) under controlled synthetic drift scenarios (3σ and 5σ). Results across 3 random seeds demonstrate: (i) an average 1.95-2.13% MAE improvement while strictly maintaining peak error at or below the static baseline, (ii) superior recovery dynamics, reducing Time-to-Recover (TTR) by 29%, and (iii) domain-adaptive behavior without hyperparameter retuning. SWS offers a transparent, low-latency baseline for operational forecasting.
Ahmad et al. (2026) studied this question.