Question on the scale-stationarity assumption in MeanScaleUniformBins and possible ways to relax it

[HarvestStars]

Hi Chronos Team,

Recently I am about to use Chronos/bolt as a codebase for my Crypto Trading Research. While I reviewed the tokenizer, I found an interesting setting in the "scale" that the tokenizer computes a mean-scale from the historical context and reuses it for both the label window and the decoding of future predictions..

This seems imply a (local) scale stationarity assumption. For non-stationary series (e.g., financial data with volatility shifts), this may introduce systematic bias (under- or over-estimated amplitudes).

Example:
When a market transitions from a sideways consolidation phase into a strong uptrend, the mean-scale estimated from the calm, low-volatility period will underestimate the amplitude of the upcoming surge. As a result, the decoded forecast may appear “too flat” and fail to capture the true strength of the move.

This seems related to the discussion in issue #57.

I’d like to confirm ask whether there are recommended ways (existing or planned) to relax it, such as sliding/learnable scaling, multi-scale tokens, or quantile normalization.
(Btw, I also noticed that issue #249 proposed a multi-scale encoder design, which seems relevant conceptually, though it doesn’t directly address the scale normalization assumption at the tokenizer level.)

[abdulfatir]

@HarvestStars The scaling here is only intended to normalize the values and make them amenable for downstream processing. There's no stationarity assumption here and non-stationary scales should not be a problem for the model.

Unrelated: have you checked out Chronos-2? It natively supports univariate, multivariate and covariate-informed forecasting tasks.

[HarvestStars]

oh, thanks @abdulfatir
yes I will check the chronos-2 paper this week