Chronos-2 Revisited: When Zero-Shot Forecasting Meets Real-World Data, Does the Magic Hold?

In a previous post I had illustrated how Chronos 2 outperformed the standard Box Jenkins and Holt Winters forecasts from Python as well the forecasts from a popular Commercial time series forecasting package.  
 
My guess is Chronos 2 has this data memorized based on all the training data sets provided. This age-old data was provided by Box Jenkins in their classic textbook on Time Series Analysis and Control published in 1970.  
 
Perhaps Chronos was trained to spit out the forecast from its memory when it was able to match all the in-sample data points……. 
 
Zero shot forecasting?  
 
Perhaps worth just zero in real-life forecasting?!  
 
The jury is still out on what it can do with all this training for the unknown future where forecasting really matters. Our initial tests on real-life “sh*tty” data from industry do not look promising with Chronos 2. More to come on this.  
 
But I am curious why memorized data that has the future numbers etched in its memory cannot help Chronos 2 to forecast perfectly?!   
 
Why not extend zero shot to come up with a Zero Mape forecast?! Just puzzled. 

Want to understand how AI and foundation models fit — or don’t yet fit — into your demand planning toolkit? Reach out to the Valtitude team for an honest assessment and explore our demand planning and forecasting services to see how we help organizations build forecasting capability that performs in the real world.