Advances in Deep Learning for Time Series Forecasting/Classification Winter 2024
Another year has passed and with it we have seen new research in the field of deep learning for time series. Although LLMs/NLP and stable diffusion techniques received all the hype throughout 2023, we have seen slow but steady progress in the time series domain. Conferences such as Neurips, ICML, and AAAI have showcased incremental improvements to the transformer architecture (BasisFormer, Crossformer, Inverted Transformer, and Patch Transformer), new architectures that synthesize numerical time series data with text and imagery (CrossVIVIT), the possible application of LLMs to time series directly (LLModels are Zero-Shot Time Series forecasters), and new forms of time series regularization/normalization techniques (SANs). There was also a fair amount of sensationalist fluff that piggybacked on the success of LLMs without any experimental or theoretical rigor (e.g. TimeGPT) that I will address at the end of the article.
Flow Forecast Framework Updates:
Before jumping into the academic papers we have some exciting new framework updates to divulge. Over the course of 2023 we added many new features and models to Flow-Forecast. Principally we have:
- Added full support for training and forecasting on multiple series for different…
