Beyond Words: Adapting NLP Methodologies for Time Series Forecasting Challenges

Although Natural Language Processing (NLP) methods are highly effective in handling textual data, their potential for handling other kinds of data, like time series, is still largely unexplored. The effectiveness of several NLP models for time series forecasting is examined in this work. We assess five models, ranging in architecture from transformers to convolutional networks: Autoformer, DLinear, WaveNet, DeepAR, and SimpleFeedForward. WaveNet exhibits superior forecasting performance with a Minimal Absolute Scaled Error (MASE) of 1.163, highlighting the usefulness of NLP techniques in time series analysis. Our results demonstrate the versatility and the efficacy of NLP techniques in identifying complex patterns in time series data. A synergistic combination of attention techniques is proposed to improve model flexibility and accuracy through additional research on sparse-dense attention mechanisms. This novel approach has the potential to open up new directions for advanced and flexible time series forecasting techniques.