T2F: Forecasting Irregularly Sampled Time Series via Time Features

Time series forecasting under irregular sampling is challenging because temporal information cannot be inferred solely from sample ordering. Although several neural architec- tures have been specifically developed for irregularly sampled data, the extent to which forecasting performance depends on model complexity rather than temporal representation remains unclear. In this work, we investigate T2F, a feature-based framework for forecasting regular and irregular time series. The forecasting task is reformulated as a supervised learning problem in which temporal information is explicitly encoded through engineered features. In particular, we compare absolute timestamp representations with local time-difference (∆t) encod- ing, which captures the temporal spacing between observations and between the last observation and the prediction target. T2F is evaluated via an MLP model on synthetic benchmarks and real-world energy consumption data under both regular and irregular sampling conditions. Experimental comparisons include naive baselines, linear models, tree-based methods, and GRU-D, a recurrent architecture designed for irregular time series. Results show that forecasting performance under irregular sampling depends strongly on temporal representation. While absolute timestamps provide useful global temporal context, local ∆t encoding consistently improves predictive accuracy, particularly in short-term forecasting. In the considered real-world irregular setting, the MLP with ∆t encoding outperforms traditional base- lines and GRU-D despite its substantially simpler architecture. Overall, the results suggest that carefully designed temporal representations can make simple feed-forward models highly competitive for moderately irregular short-term forecasting, with no specialized recurrent or continuous-time architectures.