Recursive TSFM

A small JAX/Flax time-series foundation model based on Moirai 2 that learns to forecast long horizons by rolling its transformer hidden state forward instead of autoregressing decoded values.

The code and results used for the recursive Moirai post are frozen on the blog/recursive-moirai branch.

# Prep data
uv run --script scripts/prep_eval_data.py --context_len 8192
uv run --script scripts/prep_train_data.py

# Train
uv run main.py --run_name test_run

# Evaluate
uv run gifteval.py eval_windows --run_dir runs/test_run