forecasting-reverso

Reverso Time Series Forecasting

Produce zero-shot univariate time series forecasts using the Reverso foundation model family (arXiv:2602.17634), implemented in NumPy/Numba for CPU-only container execution.

Setup (run once per conversation)

uv pip install numba --system --break-system-packages
cp /mnt/skills/user/forecasting-reverso/scripts/reverso.py /home/claude/reverso.py
cp /mnt/skills/user/forecasting-reverso/scripts/load_checkpoint.py /home/claude/load_checkpoint.py

Obtaining Weights

Two paths depending on network access:

Path A: Direct download (HuggingFace allow-listed)

import urllib.request, os
os.makedirs("/tmp/reverso", exist_ok=True)
url = "https://huggingface.co/shinfxh/reverso/resolve/main/checkpoints/reverso_small/checkpoint.pth"
urllib.request.urlretrieve(url, "/tmp/reverso/checkpoint.pth")

Path B: User upload (HuggingFace not accessible)

If the download fails with a network error, tell the user:

I can't reach HuggingFace from this environment. Please download the checkpoint from https://huggingface.co/shinfxh/reverso/blob/main/checkpoints/reverso_small/checkpoint.pth and upload it here.

Then load from /mnt/user-data/uploads/checkpoint.pth.

Loading weights

from load_checkpoint import load_checkpoint
weights = load_checkpoint("/tmp/reverso/checkpoint.pth")  # or upload path

Model Configuration

Reverso Small uses this config (matching the published args.json):

from reverso import ReversoConfig
config = ReversoConfig(d_model=64, module_list=["conv", "attn", "conv", "attn"])

Forecasting

from reverso import forecast, warmup_jit
warmup_jit()  # ~2s one-time JIT compilation

result = forecast(
    series=data,               # 1-D array/list of floats
    prediction_length=96,      # how many future steps
    weights=weights,           # dict from load_checkpoint
    config=config,
)

The function handles preprocessing (NaN interpolation, padding, min-max normalization) and autoregressive rollout internally.