Transformes for Time Series: Is the New State of the Art (SOA) Approaching?
summary
Claims that transformers beat LSTMs on long-horizon forecasting for time series get tested with open source Informer and SpaceTimeFormer code, using prob-sparse attention and timestamp/holiday embeddings. A microservice latency use case uses ~60 metrics expanded to ~250 input dims and ~5k timestamps (8k train / 2k test split) to predict front-end latency; Informer achieves ~0.6 MSE on a 36-step forecast while LSTM is notably worse, but transformers incur higher training and inference time.
AI-written summary · about
chapters (5)
- 0:00Start
- 1:41Transformers Family Tree
- 13:21What is a transformer? 4/4 DECODER
- 18:53Neural networks approach
- 22:58TRANSFORMERS - How do they help?
related talks
31mAdvanced machine learning time series....
31mFoundational Time Series Models in Practice: The Future of Forecasting, or Just Hype?
21mFoundational Models for Time Series Forecasting: are we there yet?
31mFoundational Models for Time Series Forecasting: are we there yet?
33mFrontiers of Foundation Models for Time Series
19mForecasting for Everyone - Foundation Models and the Future of Time Series Prediction
31mML time series forecasting methods
30mFoundational Models for Time Series Forecasting: are we there yet