How Simple Can Multi-Horizon Forecasting Be?

Multi-horizon time-series forecasting requires models to capture heterogeneous temporal dynamics that influence predictions differently across short-, medium-, and long-term horizons. Existing forecasting models often process these dynamics using a uniform modelling strategy or increasingly complex transformer and attention architectures designed to capture dependencies across different temporal scales. However, it remains unclear whether such architectural complexity is necessary to maintain accurate predictions as the forecast horizon increases. To investigates this simple question: how simple can a forecasting model be while remaining effective across short-, medium-, and long-term horizons? We introduce D-LML (Decomposition–Linear–MLP–Linear), a lightweight architecture for direct multi-horizon forecasting for time series data. D-LML decomposes each historical input window into trend, seasonal, and residual components and processes them using two linear mappings and a compact shallow multilayer perceptron before additively reconstructing the future sequence. We evaluate D-LML across multiple datasets and progressively increasing prediction horizons against established forecasting architectures representing linear, MLP-based, recurrent, multiscale, and attention-based modelling paradigms. Our experiments showed that D-LML achieves competitive state-of-the-art performances in short, medium and long term across several commonly used datasets, with the superior forecasting accuracy while requiring substantially fewer trainable parameters than high-capacity neural baselines. These findings suggest that carefully structured lightweight models provide a strong alternative for accurate and computationally efficient multi-horizon time-series forecasting.