Adapting State-of-the-Art Time Series Models for Resource-Constrained Deployment in Africa

Long-term time series forecasting has trended toward increasingly complex Transformer architectures, yet recent work shows simple MLP-based models can match or exceed them, though even these may retain unnecessary complexity. In much of Africa, where GPU access is limited, minimizing model complexity without sacrificing performance is a practical necessity. Using the SOFTS architecture as a case study, we show that even simplified MLP models can be simplified further: removing SOFTS's dense inter-channel modeling module yields a variant that matches or improves forecasting accuracy on low-channel ETT benchmarks (16 first-place MSE finishes vs. 5 for the original), while CPU-only hardware testing shows up to 81.7\% lower latency jitter, 50\% higher throughput, and reduced energy consumption. However, on larger-channel datasets, the simplified variant underperforms SOFTS, confirming that channel mixing matters as channel count scales. We argue practitioners in resource-constrained environments should treat state-of-the-art architectures as modular baselines rather than rigid monoliths, making targeted simplification a practical path toward sustainable AI deployment in Africa.