| --- |
| tags: |
| - time-series |
| - forecasting |
| - ensemble |
| - agentic |
| - gift-eval |
| --- |
| |
| # Cobra Agent |
|
|
| Cobra Agent is an agentic forecasting ensemble system developed by Dalpha. |
|
|
| ## Model Details |
|
|
| Cobra Agent combines the outputs of multiple time-series foundation models instead of relying on a single forecaster. The system analyzes the target time series, dataset configuration, and forecast behavior of candidate models, then dynamically constructs a weighted ensemble forecast. |
|
|
| The current version uses the following pretrained forecasting models: |
|
|
| - [Chronos-2](https://huggingface.co/amazon/chronos-2) |
| - [TimesFM-2.5-200M PyTorch](https://huggingface.co/google/timesfm-2.5-200m-pytorch) |
| - [Toto-2.0-2.5B-FT](https://huggingface.co/Datadog/Toto-2.0-2.5B-FT) |
| - [TiRex-1.1 GIFT-Eval](https://huggingface.co/NX-AI/TiRex-1.1-gifteval) |
|
|
| Cobra Agent is designed to improve robustness across heterogeneous forecasting tasks by adapting ensemble weights according to task-level and forecast-level signals. |
|
|
| ## Method Overview |
|
|
| Cobra Agent estimates the expected relative reliability of each candidate model forecast for a given prediction task. Instead of selecting a single model or relying on fixed ensemble weights, the system evaluates candidate forecasts in a model-conditional manner and converts the estimated reliability scores into adaptive ensemble weights. |
|
|
| The system uses information from the input time series, dataset configuration, model identity, and forecast-level patterns to estimate how suitable each foundation model forecast is for the current task. Some of the time-series input features were inspired by the feature design used in Toto-FnF, while Cobra Agent extends this idea with model-conditional score estimation and forecast-level fingerprint signals. |
|
|
| It also incorporates validation-derived priors and conservative weighting controls to reduce over-reliance on any single model or heuristic. |
|
|
| ## Evaluation Policy |
|
|
| Cobra Agent was developed without using GIFT-Eval test labels, test metrics, or leaderboard feedback for training, model selection, or ensemble-weight tuning. |
|
|
| All routing and weighting decisions were derived from validation forecasts and validation metrics only. Test forecasts were generated after the validation-selected system was fixed and were used only to produce the submitted GIFT-Eval result file. |
|
|
| ## State of Research |
|
|
| Cobra Agent is an internal research system from Dalpha. |
|
|
| A detailed technical report and reproduction materials are planned for a future public release. |
|
|
