Time series and forecasting · insight-1253
Handle sparse and flat series before sending them to a mean-scaled, tokenizing forecaster
- Evidence type
- documented fact
- Difficulty
- hard
- Roles
- Data scientist, ML engineer, Applied AI
- Industries
- manufacturing, retail, logistics
Summary: situation, action, reason, boundary
Situation
A pretrained model that scales each series by the mean absolute value of its context is run on spare-part demand where one week in twenty has a sale of one unit.
Action
Route sparse series to intermittent-demand methods or aggregate to a coarser bucket until nonzero values are frequent. Handle all-zero contexts explicitly and report the intermittent segment separately.
Reason
Mean scaling divides by the mean absolute context value, and the tokenizer covers a bounded range, so spikes of a unit every n steps cannot be represented once n exceeds the range limit.
Boundary
Zero-shot tokenization-based forecasters on count, sparse or near-constant series.
Symptom or risk
The scale becomes small, so the rare sale maps to a large scaled number outside the token range, and an all-zero window has a scale of zero.
Exceptions
Series with frequent non-zero values, or other normalisation schemes, avoid this failure. It is described for the original token-based Chronos models; patch-based releases handle inputs differently, so check the model you use.
Trade-offs
Coarser buckets lose timing information that a stocking decision may need.
Pitfalls
Averaging away the failure on the sparse segment in a mixed-portfolio benchmark.
Example
Hypothetical: one unit sold every 20 weeks gives mean absolute value 1/20 = 0.05, so a sale scales to 20, beyond the +15 bin edge; the largest representable demand is 15 x 0.05 = 0.75 units.
Lessons and concepts
Claims and their sources
Chronos scales each series by the mean of its absolute context values, quantizes into bins over a bounded range of about minus 15 to plus 15, and the paper shows that sparse series with unit spikes every n steps exceed the representable range once n is large.