Long-term Forecasting with TiDE: Time-series Dense Encoder

Authors: Abhimanyu Das, Weihao Kong, Andrew Leach, Shaan K Mathur, Rajat Sen, Rose Yu

TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we present our main experimental results on popular long-term forecasting benchmarks. We also perform an ablation study that shows the usefulness of the temporal decoder. We present Mean Squared Error (MSE) and Mean Absolute Error (MSE) for all datasets and methods in Table 2.
Researcher Affiliation Collaboration Abhimanyu Das EMAIL Google Research; Weihao Kong EMAIL Google Research; Andrew Leach EMAIL Google Cloud; Shaan Mathur EMAIL Google Cloud; Rajat Sen EMAIL Google Research; Rose Yu EMAIL University of California, San Diego
Pseudocode No The paper describes the model architecture and operations in text and through diagrams (Figure 1), but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our model is trained in TensorFlow (Abadi, 2016) and we optimize using the default settings of the Adam optimizer (Kingma and Ba, 2014). We provide our implementation in the supplementary with scripts to reproduce the results in Table 2.
Open Datasets Yes We use seven commonly used long-term forecasting benchmark datasets: Weather, Traffic, Electricity and 4 ETT datasets (ETTh1, ETTh2, ETTm1, ETTm2). We refer the reader to (Wu et al., 2021) for a detailed discussion on the datasets. In order to showcase our model s ability to handle static attrubutes and complex dynamic covariates we use the M5 forecasting competition benchmarks (Makridakis et al., 2022).
Dataset Splits Yes In all the datasets, the train:validation:test ratio is 7:1:2 as dictated by prior work.
Hardware Specification Yes All the experiments in this section were performed using a single NVIDIA T4 GPU on the same machine with 64 core Intel(R) Xeon(R) CPU @ 2.30GHz.
Software Dependencies No The paper mentions that the model is trained in Tensorflow and uses the Adam optimizer, citing their respective foundational papers. However, it does not provide specific version numbers for Tensorflow or any other software libraries used, which is required for reproducibility.
Experiment Setup Yes We tune our hyper-parameters using the validation set rolling validation error. We provide details about our hyper-parameters in Appendix B.3. As global dynamic covariates, we use simple time-derived features like minute of the hour, hour of the day, day of the week etc which are normalized similar to (Alexandrov et al., 2020)... We use the validation loss to tune the hyper-parameters per dataset. We report the specific hyper-parameters chosen for each dataset in Table 8.