finsignals

Numerai Signals from FinField P2P field data — FinFacts in, a submit-ready bloomberg_ticker,signal CSV out.

• Identity ticker map — Numerai keys its universe on bloomberg-style composite tickers ("AAPL US"), the exact shape FinField uses, so no symbol translation layer. The live universe file is supplied by you (fetching it needs a Numerai account); finsignals is offline-first and hard-codes no volatile URLs.
• Exact features — momentum (63d/252d) and volatility from finfield:close facts, all Decimal at fixed precision (no floats, ever): same facts, same feature bytes on every node.
• Features are facts — feature_facts mints finsignals:momentum_63d etc. as scale-6 FinFacts whose derived_from lists every close CID used, so signals knit back into the field and, being continuous quantities, are votable via vank weighted-median consensus (FinField/knit).
• Consensus-first — when the field disagrees on a price, feed the finfield-consensus fact (from finknit.vote.float_consensus) rather than any single observer's close.
• Valid by construction — rank r of n maps to r/(n+1), strictly inside (0,1); ties break on ticker sort; the CSV is byte-identical across runs.

Pure Python, stdlib only.

pip install "finsignals @ git+https://github.com/FinField/signals"

End-to-end: field facts → submission.csv

from finsignals import (feature_facts, fundamentals, compose_score,
                        rank_signal, write_submission, load_universe, intersect)

tickers = intersect(field_entities, load_universe("numerai_universe.csv"))

scores = {}
for ticker in tickers:
    facts = field_facts_for(ticker)          # prices + fundamentals from the field
    feats = feature_facts(f"ticker:{ticker}", facts, asof_day="2026-07-03")
    scores[ticker] = compose_score({f.concept: f.decimal for f in feats})
    # optionally weave `feats` back into the knitweb so peers can vote on them

write_submission("submission.csv", rank_signal(scores))
# bloomberg_ticker,signal — weekly eras key on a Friday date; upload as-is

The score recipe is fixed and documented: momentum_252d - momentum_63d when both exist, else whichever momentum is available, else a neutral 0.

Real-data example: value factor from the live feed

examples/feed_value_factor.py goes from the signed FinField/feed corpus to a submit-ready file in one command — no market-data vendor:

git clone https://github.com/FinField/feed /tmp/finfield-feed
python examples/feed_value_factor.py --feed /tmp/finfield-feed/feed --out submission.csv
# decoded 87313 facts for 6500 entities (0 rejected)
# wrote submission.csv with 4182 tickers

The signal is the Fama-French value factor — book equity over free-float market cap (finfield:book_to_float_mcap) — cross-sectionally ranked into (0,1). Byte-identical across runs.

Part of the FinField field: facts · scrapers · knit · agents · crypto

Extended factor zoo — the edge features

finsignals.factors adds the academically-validated cross-sectional factors that actually move Signals correlation, all computed from the same finfield:close facts (no new data source), all exact-Decimal and mintable back into the field via factor_facts (scale-6 pure, derived_from CID chains, votable):

factor	concept	why it has edge
12-1 skip momentum	finsignals:skipmom_252_21	residual momentum, strips short-term reversal
multi-horizon momentum	finsignals:momentum_21d / _126d	captures medium-term trend
short-term reversal	finsignals:reversal_5d	1-week mean reversion
downside volatility	finsignals:downside_vol_63d	semideviation / downside risk premium
lottery (MAX)	finsignals:max_return_21d	high-max stocks under-perform (negative predictor)
52-week-high proximity	finsignals:high_252d_ratio	George–Hwang nearness-to-high
moving-average gap	finsignals:ma_gap_50 / _200	trend / distance from MA
return skewness	finsignals:skew_63d	idiosyncratic-skew premium
long-horizon volatility	finsignals:volatility_252d	risk normalization

from finsignals import factor_facts
feats = factor_facts(f"ticker:{ticker}", field_facts_for(ticker), asof_day="2026-07-03")
# each fact is finsignals:<factor> at scale 6, derived_from the exact close CIDs used

The standard feature_facts pack (momentum_63d / momentum_252d / volatility_63d) is unchanged and byte-stable; factor_facts is purely additive. Build your model on the full factor matrix, or weave the facts back so peers can vote on them.