Glimpse of dataset - (1) synthetic time series generation

Several studies have trained models using synthetic time series data as a complement to real-world data or even as a standalone approach, showing comparable zero-shot performance. 

(ForecastPFN, TimsFM, Fu et al., 2024)

 

Given my limited resource budget, incorporating synthetic data into my training dataset is a viable option. 

 

Additionally, conducting an ablation study to evaluate the effectiveness of synthetic data would also be interesting.

The important question is, given the constraints on time and resources, what is the optimal mixture of training data that will fit within my budget? This mixture should ensure that training is not overly burdensome while still providing the model with enough information to learn time series patterns so enabling the model to accurately extrapolate future trajectories.

---

I generated three frequency groups (daily, weekly, and monthly) of 10,000 time series with a context length of 2,048. Each has distinct trends and seasonal patterns.

The generation formula I used is based on the ForecastPFN paper, but I did not include a noise term, as TimesFM also did not include any noise term.