|
--- |
|
library_name: transformers |
|
base_model: |
|
- amazon/chronos-t5-small |
|
pipeline_tag: time-series-forecasting |
|
--- |
|
# Model Card for Chronos T5 Small Fine-Tuned Model |
|
|
|
|
|
<img align="center" height="350" src="https://media3.giphy.com/media/v1.Y2lkPTc5MGI3NjExYjRmcWUwaGFkbW1lczJoYzBjbHBxZjMyeDdhdDQycGdzamwyOGhiZyZlcD12MV9pbnRlcm5hbF9naWZfYnlfaWQmY3Q9Zw/ZtB2l3jHiJsFa/giphy.gif"/> </p> |
|
|
|
|
|
## Summary |
|
|
|
This model is fine-tuned for time-series forecasting tasks and serves as a tool for both practical predictions and research into time-series modeling. It is based on the `amazon/chronos-t5-small` architecture and has been adapted using a dataset with 15 million rows of proprietary time-series data. Due to confidentiality restrictions, dataset details cannot be shared. |
|
|
|
## Fine-Tuning Dataset |
|
|
|
The model was fine-tuned on a proprietary dataset containing 15 million rows of time-series data. While details about the dataset are confidential, the following general characteristics are provided: |
|
- The dataset consists of multi-dimensional time-series data. |
|
- Features include historical values, contextual attributes, and external covariates relevant to forecasting. |
|
- The data spans multiple domains, enabling generalization across a wide range of forecasting tasks. |
|
|
|
This large-scale dataset ensures the model captures complex patterns and temporal dependencies necessary for accurate forecasting. |
|
|
|
## Evaluation |
|
|
|
### Testing Data, Factors & Metrics |
|
|
|
#### Testing Data |
|
|
|
The model was evaluated using several publicly available time-series datasets, including: |
|
- **electricity_15min** |
|
- **monash_electricity_hourly** |
|
- **monash_electricity_weekly** |
|
- **monash_kdd_cup_2018** |
|
- **monash_pedestrian_counts** |
|
|
|
#### Factors |
|
|
|
Evaluation was conducted across datasets representing various domains such as electricity usage, pedestrian counts, and competition data. |
|
|
|
#### Metrics |
|
|
|
Two primary metrics were used for evaluation: |
|
- **MASE (Mean Absolute Scaled Error):** A normalized metric for assessing forecast accuracy. |
|
- **WQL (Weighted Quantile Loss):** Measures the quality of probabilistic predictions. |
|
|
|
### Results |
|
|
|
| Dataset | Model | MASE | WQL | |
|
|-----------------------------|------------------------------|--------|---------| |
|
| electricity_15min | amazon/chronos-t5-small | 0.425 | 0.085 | |
|
| monash_electricity_hourly | amazon/chronos-t5-small | 1.537 | 0.110 | |
|
| monash_electricity_weekly | amazon/chronos-t5-small | 1.943 | 0.086 | |
|
| monash_kdd_cup_2018 | amazon/chronos-t5-small | 0.693 | 0.309 | |
|
| monash_pedestrian_counts | amazon/chronos-t5-small | 0.308 | 0.247 | |
|
|
|
#### Summary |
|
|
|
The fine-tuned model performs well on short-term electricity datasets (e.g., **electricity_15min**) with low MASE and WQL values. Performance varies depending on the dataset's characteristics, particularly with longer-term or aggregated data. |
|
|
|
## Technical Specifications |
|
|
|
### Model Architecture and Objective |
|
|
|
The model is based on the `amazon/chronos-t5-small` architecture, fine-tuned specifically for time-series forecasting tasks. It leverages pre-trained capabilities for sequence-to-sequence modeling, adapted to handle multi-horizon forecasting scenarios. |
|
|
|
## Citation |
|
|
|
If you use this model in your research or applications, please cite it as: |
|
|
|
## Contact: |
|
|
|
[NIEXCHE (Fevzi KILAS)](https://niexche.github.io/) |
|
|
|
 |
|
|
|
|
