idun-numerai

Numerai on top of idun's public interface: Numerai's data in idun's packed format, the walk-forward refits, the labels, Numerai's scores in NumPy, fixed presets for each learner, and model uploads.

pip install idun-numerai    # Python 3.12–3.14; idun's Metal backend needs Apple silicon with macOS 15 or later

Train and score

from idun import gbm
from idun.data import PackedDataset
from idun_numerai import presets
from idun_numerai.data import build_cache
from idun_numerai.download import download
from idun_numerai.score import corr, per_era

download("v6")  # Numerai's v6 training and validation files and the feature list (public data, no credentials)
build_cache("v6", "cache", seed=0)  # streamed into a memory-mapped packed dataset; never the whole panel in memory

data = PackedDataset.open("cache", binning="r2")  # the last walk-forward window's binning
window = data.refits[2]
model = gbm.train(presets.GBM_C, data, rows=slice(0, window.train_hi), backend="metal")
rows = slice(window.pred_lo, window.pred_hi)
scores = per_era(corr, data.era[rows], model.predict(data, rows=rows, backend="metal"), data.target[rows])

• idun_numerai.download, idun_numerai.data.build_cache: the data, checked and packed (4-bit codes, the named factors binned per window, the country as a categorical column).
• idun_numerai.refits: the walk-forward windows (training rows, a purge, the predicted rows).
• idun_numerai.inputs: the labels: classes rint(4 · y) for the networks' classification, per-era z for regression.
• idun_numerai.presets: fixed configurations of idun_gbm (GBM_C and its seed bag), TabM and RealMLP.
• idun_numerai.score: NCORR and CORR per era in NumPy and SciPy, as numerai-tools defines them (checked against it per era in the tests; numerai-tools is not a dependency).

Model uploads

Numerai's model uploads run a pickled predict(live_features, live_benchmark_models) on its servers, where idun is not installed. idun_numerai.upload carries trained models as arrays and predicts with NumPy alone:

from idun_numerai import upload

predict = upload.predictor([model, [tabm_s0, tabm_s1], realmlp], data, weights=[2, 1, 1])
upload.save(predict, "model.pkl")  # run under Python 3.12, the runtime's

idun_gbm's upload predictions equal Booster.predict's bit for bit; TabM's and RealMLP's are their float64 reference backend's. A list is a bag (its mean); several entries are blended by rank.

Licence

Apache-2.0.