Nikolay Laptev, Jiafan Yu, Ram Rajagopal

Feb 15, 2018 (modified: Feb 15, 2018) ICLR 2018 Conference Blind Submission readers: everyone Show Bibtex
  • Abstract: Reliable and accurate time-series forecasting is critical in many fields including energy, finance, and manufacturing. Many time-series tasks, however, suffer from a limited amount of training data (i.e., the cold start problem) resulting in poor forecasting performance. Recently, convolutional neural networks (CNNs) have shown outstanding image classification performance even on tasks with small-scale training sets. The performance can be attributed to transfer learning through CNNs’ ability to learn rich mid-level image representations. However, no prior work exists on general transfer learning for time-series forecasting. In this paper, motivated by recent success of transfer learning in CNN model and image-related tasks, we for the first time show how time-series representations learned with Long Short Term Memory (LSTM) on large-scale datasets can be efficiently transferred to other time-series forecasting tasks with limited amount of training data. We also validate that despite differences in time-series statistics and tasks in the datasets, the transferred representation leads to significantly improved forecasting results outperforming majority of the best time-series methods on the public M3 and other datasets. Our online universal forecasting tool, DeepCast, will leverage transfer learning to provide accurate forecasts for a diverse set of time series where classical methods were computationally infeasible or inapplicable due to short training history.