TugaYousif/Falcon-TST_challenging_cases

Falcon-TST Forecasting Experiment

Model

Model used in this experiment:
Falcon-TST Large

Full Code

The full experiment code can be found here:
https://colab.research.google.com/drive/193rraeQxrkGQIgh_Ab4Sp4939Pa0oEAu?usp=sharing

Loading the Model

Example snippet used to load the model in Google Colab:

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained(
    "ant-intl/Falcon-TST_Large",
    trust_remote_code=True
)

Model Overview and Performance

Falcon-TST is a transformer-based foundation model designed for time series forecasting.
It is trained on large collections of time series data and can perform forecasting across many domains without task-specific training.

In the experiments conducted for this dataset, Falcon-TST demonstrated competitive performance with state-of-the-art time-series foundation models, including:

• TimesFM 2.5 developed by Google
• Chronos-T5 developed by Amazon

Across the evaluated datasets:

• Falcon generally outperformed Chronos-T5 in most forecasting tasks.
• However, TimesFM 2.5 consistently produced the best overall accuracy.
• Falcon achieved particularly strong performance on the weather dataset, suggesting it handles smoother seasonal patterns well.

Overall, Falcon-TST provides strong baseline performance comparable to modern foundation models, though some limitations were observed in specific dataset characteristics.

Model Blind Spots

Several patterns emerged where Falcon-TST performed poorly:

• Small dataset size (datasets with only a few hundred observations)
• Extreme amplitude variation where the scale of values changes dramatically
• Non-stationary structures where the statistical properties of the series shift over time

These characteristics appear to reduce the model's ability to accurately capture future dynamics.

Suggested Fine-Tuning Data

To improve the model's performance on these blind spots, fine-tuning should focus on datasets that include:

• Small and medium-sized time series
• Time series with large variance changes
• Data containing structural breaks or regime shifts
• Series with extreme spikes or bursts
• Non-stationary real-world signals

Including such datasets would help the model learn patterns that are underrepresented in its original training distribution.

How to Assemble Such a Dataset

A suitable dataset could be assembled by combining time series from multiple real-world domains:

• weather and climate observations
• air quality measurements
• traffic flow data
• financial time series
• energy consumption datasets
• retail demand series

Public repositories such as Kaggle, UCI Machine Learning Repository, and existing forecasting benchmarks can provide diverse time series with the desired characteristics.

The dataset should include series with:

• varying lengths
• different scales
• different levels of noise
• irregular or changing seasonal patterns

Estimated Dataset Size for Fine-Tuning

To effectively address these blind spots, a fine-tuning dataset would likely need:

• tens of thousands of time series
• covering multiple domains
• with variable lengths and structures

A dataset on the order of 50k–100k time series would likely provide sufficient diversity for the model to learn robust representations of small, noisy, and non-stationary signals.