nkh/timerouter-v1

TimeRouter

Trained XGBoost router weights for TimeRouter, the GIFT-EVAL submission that routes among 4 frozen time-series foundation models (Chronos-2, FlowState, PatchTST-FM, Sundial) with a margin/diversity gate and a CV-inverse-weighted fallback.

LB MASE = 0.6746 on the full 97-config GIFT-EVAL test suite.

Files

File	Description
seed42.json … seed46.json	5-seed XGBoost OvA ensemble. 305-dim features, 400 trees × depth 8 × lr 0.05 × subsample 0.8, random_state ∈ {42..46}, tree_method="hist".
pool_metadata.json	Pool config ({chronos, flowstate, patchtst_fm, sundial}), 305-feature column order, and gate thresholds (tau_m, tau_d) = (0.15, 0.02).

Usage

These are the checkpoints loaded by gift_eval/run_eval.py --ckpt-dir <this folder> in the TimeRouter repository; see its README for the full two-environment setup and run instructions. Requires xgboost >= 2.x.

from huggingface_hub import snapshot_download
ckpt_dir = snapshot_download("nkh/timerouter-v1")
# then: python gift_eval/run_eval.py --ckpt-dir $ckpt_dir ...

Citation

If you use these checkpoints, please cite the TimeRouter repository (UConn Data Science and Intelligent System (DSIS) Research Lab).