Skip to content

David v Goliath: re-run as foundation models improve #97

Description

@microprediction

The page at https://skaters.microprediction.org/timesfm.html currently concludes that TimesFM (2.5, 200M) only beats Laplace on degenerate repeated-value series, and that Laplace wins 69/69 on likelihood over the continuous split with a 2.26-nat median gap.

Foundation models improve quickly and this verdict should not be assumed durable. Re-run benchmarks/foundation_study.py when any of these change:

  • a new TimesFM checkpoint or size (2.5 → next)
  • native density or improved quantile heads (the LL comparison is currently read through quantile reconstruction)
  • long-context variants (only 256 tested)
  • the unrun cells: MAE of the median, the levels protocol

The page promises to be updated if any rerun says otherwise.

🤖 Generated with Claude Code

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions