bolt vs t5

[doninocode]

Hi,

1. When passed the same input timeseries why does the t5 model produce a forecast with a realistic variance whereas the bolt model produces a smooth forecast with very minimal oscillations?

2. When using the t5 model, given the same input timeseries, I get vastly different results every time I run the model, anything I can do achieve a consistent result? I tried setting a seed, which worked, but it seems a very random way to get the consistency I'm looking for.

T5 example:

Bolt example: with same input series

Thank you!

[lostella]

For 1, what forecast is being plotted in the two cases? Are you invoking predict and plotting (one slice of) its output tensor? If that’s the case, then you’re plotting one sample trajectory from chronos-t5 (which has the same random-walk-like behavior as historical data) and one quantile for chronos-bolt (this model directly produces quantile estimates, no samples).

For 2, chronos-t5 is inherently sampling based, so as long as you observe one output sample, fixing the seed is the only way to make the model deterministic. If instead you want to look more at the “full” predictive distribution, looking at multiple samples (or quantiles) is the way to go: increasing num_samples (default 20) will make the predictive distribution more consistent from on run to the next (but not identical: for that, fixing the seed is the only way).

Edit: you can also set temperature=0 in chronos-t5 models, which should result in the “most likely” sample. But then again, that hardly what one wants (you get no clue about how certain/uncertain the model is)

[doninocode]

Thank you very much for explaining that. Makes more sense now.