cs-labs/khou-tabpfn-models

KHOU TabPFN bundle

Trained model + features + WU history for daily-high temperature prediction at Houston Hobby (KHOU, US). Built by cold-path/3-weather-onboard-station/.

This station has 4 trade slot(s): c12_d1, c18_d1, c00_d0, c06_d0. Each slot is a (cycle, init_day_offset) pair derived from the station's UTC offset relative to the local trade window (D-1 06:00 -> D 10:00). See config.py for the per-slot fire times + n_rollouts.

Layout

KHOU-tabpfn-models/
├── config.py                       per-station constants + slot list
├── build_features.py               feature row builder (slot-aware)
├── inference.py                    TabPFN in-context predictor
├── requirements.txt
├── data/
│   ├── wu_raw.parquet              ~hourly WU obs, 2005-present
│   └── wu_ground_truth.parquet     daily highs derived from raw obs
├── models/
│   └── <slot_id>/training_data.parquet  per-slot TabPFN context
│                                        (4 subdirs)
└── backtesting/
    ├── README.md                   date windows
    ├── polymarket_history.csv      one row per (date, slot, bucket)
    └── <slot_id>_oof.parquet       8-fold OOF preds on the holdout

Backtesting

The backtesting/ folder ships out-of-sample TabPFN predictions and Polymarket history for the latest backtest window. See backtesting/README.md for the date ranges.

Quick start

from inference import predict
from config import SLOT_BY_ID

# Pick any slot from the bundle's SLOT_BY_ID:
#   `c12_d1`, `c18_d1`, `c00_d0`, `c06_d0`
out = predict(target_date="2026-05-15",
              slot_id="c06_d0",
              trajectory_nc_path="<cropped FuXi NC path>")
print(out)  # {"p10_c": ..., "p50_c": ..., "p90_c": ...}