Their stack it's in python, regularly maintained, consistent data structures and design patterns across libraries, which reduces engineering/implementation time, but also a wide range of classical methods for time series analysis, uncertainty quantification, along with novel developments , especially in the deep learning domain. The community is highly supportive, not only the engineering team from Nixtla, and most importantly, OPEN SOURCE.

From Nixtla’s ecosystem, I particularly value TimeGPT, which I used in a project focused on predicting kNDVI (Normalized Difference Vegetation Index) across 214,351 time series derived from a geospatial datacube. The model demonstrated adaptability; it effectively captured vegetation dynamics when provided with sufficient historical context. With fine-tuning and the inclusion of exogenous variables, TimeGPT achieved reliable long-horizon forecasts, outperforming traditional models across multiple metrics while better reproducing the shape, amplitude, and temporal dynamics of the kNDVI signal. TimeGPT also shows strong potential for gap-filling and temporal interpolation in satellite-derived time series, where missing observations are frequent due to cloud cover or sensor limitations. When given full historical context, it generalizes the underlying temporal patterns remarkably well, making it a valuable tool for Earth observation and remote sensing applications, provided adequate computational resources and contextual information are available.

Very easy to use, and fast. I was able to set it up quickly in my forecasting pipeline.