Fine-tune Toto for turbocharged forecasts

Datadog's release of Toto, a state-of-the-art time series foundation model, emphasizes its strong zero-shot forecasting capabilities, which have been well-received with over nine million downloads. However, in real-world applications, the availability of historical data and additional predictive features can significantly enhance forecast accuracy. To address this, Datadog has introduced fine-tuning capabilities and support for exogenous covariates, allowing users to adapt Toto to specific workloads and incorporate known future predictors, such as billing cycles or marketing campaigns, to improve forecast precision. These enhancements, demonstrated through cloud cost forecasting and public benchmark datasets, show that fine-tuning with exogenous covariates notably improves performance across various metrics. The new features are available in the Toto repository on GitHub, and Datadog encourages users to experiment with different settings to optimize forecasting accuracy for domain-specific needs.