| dc.description.abstract | Multi-horizon time series forecasting poses fundamental challenges to machine learning and statistics with applications in many domains. In direct multi-horizon forecasting, the standard approach of neural nets is to either have one output node per horizon or use a sequence to sequence method to achieve solid forecasts. This thesis proposes a novel multi-horizon forecasting scheme that only uses one output node for all horizons. The method achieves differentiation of horizons by encoding and injection of horizon metadata into the models. Furthermore, we introduce a multi-horizon time series adaptation of the Vision Transformer. Moreover, we propose three different ways in which to inject the horizon metadata for the transformer structure, yielding rich representations per horizons and improved results compared to a multilayered perceptron baseline. In addition, we provide six different ways to encode the different horizons into metadata. Lastly, we show that the correct encoding structure for the horizon metadata allow the encoding of the time series dynamics into the model. Ultimately, this allows the models to perform interpolation tasks. | |
