Secondly, we tackle the issue of varying dimensionality using our proposed Any-variate Attention mechanism. This approach simultaneously considers both the time and variate axes as a single sequence, leveraging Rotary (RoPE) and learned binary attention biases to encode the time and variate axes, respectively. Importantly, Any-variate Attention enables the model to accept an arbitrary number of variates as input.
Thirdly, we overcome the challenge of requiring flexible predictive distributions by introducing a mixture of parametric distributions. By optimizing the negative log-likelihood of a afghanistan phone number list flexible distribution, we ensure that our model is competitive with target metric optimization, a powerful feature for pre-training universal forecasters. This approach allows for subsequent evaluation using any target metric.
Lastly, to facilitate the training of our large time series model, we introduce the LOTSA, the largest collection of open time series datasets by collating publicly available sources of time series datasets. This effort aims to cover a broad spectrum of domains, consolidating datasets from diverse sources with varying formats. The resulting collection spans nine domains, with a total of 27B observations, with key statistics in Tables 2 and 3. More details on the key properties of these datasets, like the domain, frequency, number of time series, number of target variates, number.
- Board index
- All times are UTC
- Delete cookies
- Contact us