The "model" that wins is a zero-parameter 3-factor linear composite. This card documents that champion, the ML baselines that fail to beat it, and the evaluation ruler that makes the comparison honest. Code is Apache-2.0; the price data it trains on is not redistributed (see the Dataset Card).

Model summary


Task	Cross-sectional P(trade succeeds) / tradeability ranking on B3 equities
Champion	Linear fundamental tilt — equal-weight mean of cross-section ranks of `earnings_yield`, `book_to_price`, `profit_margin`
Trainable params	0 (fixed equal-weight composite, no tuning)
Construction	long top tercile, short bottom tercile, equal-weight, market-neutral
Rebalance	quarterly (63 trading days) — turnover is the whole game
Costs	square-root market impact (Tóth et al.) + 5bps half-spread, one-way
Baselines	LightGBM LambdaRank (40 feat), TabPFN v2 (foundation tabular small-N)

Why a linear composite is the "model"

The central finding is an honest invalidation: across the AFML levers the economic edge (DSR) sat at 0 for every price-derived configuration. The only configuration that crosses breakeven net of costs is a pre-registered, fixed, no-tuning fundamental tilt. SOTA ML does not improve on it:

LightGBM LambdaRank (40 features, quarterly groups) overfits: Rank IC OOS −0.014, net ~0, effective sample ~58 quarters. ~28-item queries give listwise ranking almost no signal.
TabPFN v2 (foundation model designed for small-N) ties the linear composite on the fundamental factors (+0.183 vs +0.18) and loses on the full feature set (turnover 1.10, churn-driven cost blowup).

Verdict: the bottleneck is information + cost, not model capacity. A zero-parameter linear composite equals the SOTA foundation model.

Real results (do not invent, do not inflate)

Champion — fundamental tilt (pre-registered, one-shot)

Metric	Value
Sharpe gross	+0.19
Sharpe net (one-way costs)	+0.18
PSR(net>0)	0.92
DSR (deflated, 18 trials)	0.72
Annualized (×√(252/63))	~0.35
Verdict	first & only economically viable config

Baselines

Model	Features	Turnover/rebal	Rank IC OOS	Net Sharpe	Verdict
Linear composite (champion)	3 fundamental	low (~0.4)	—	+0.18	viable
TabPFN v2	3 fundamental	low	—	+0.183	ties champion
TabPFN v2	all (TS+micro+regime+fund)	1.10	—	worse	loses (churn)
LightGBM LambdaRank	40	high	−0.014	~0	overfits

Price-signal context (why fundamentals win)

The price-derived signal (momentum/vol/fracdiff + Amihud + CUSUM/SADF regime) is statistically real (gross DSR ~0.41) but net −1.1 rebalancing every 10 days — sub-economic. Cost is turnover-driven: same ruler, opposite verdict for high- vs low-turnover. Price high-turnover gross +0.06 → net −1.1. What changes everything is turnover, not signal strength.

Evaluation ruler

The number that counts is net DSR under the deflated ruler, never gross Sharpe or Rank IC alone:

CPCV — 6 groups, 2 test → 15 splits / 5 paths; purge + embargo in trading days.
Deflated / Probabilistic Sharpe (Bailey–LdP) with effective N (unique dates / HORIZON × avg uniqueness) and a trials log deflating by all ~21 attempts (cpcv.append_trial / deflated_sharpe(..., n_trials=real_count)).
Sample uniqueness weights (AFML ch.4) on the LightGBM fit.
Costs: square-root impact, one-way convention.
Pre-registration: the fundamental tilt was registered (hypothesis + fixed construction + success criterion net Sharpe > 0) before running — no metric shopping. See fundamental_tilt.py docstring.

Intended use & limitations

Intended use: methodology reference and reproducible baseline for cost-aware, deflation-honest cross-sectional equity research. A teaching artifact on anti-self-deception evaluation.

Out of scope: live trading. Numbers are research estimates under a conservative cost floor (impact is more convex in low-ADV B3 mid-caps than √ captures).

Known limitations

Survivorship: current liquid composition, not full point-in-time membership (partially mitigated by including post-2018 IPOs; full PIT membership pending).
Small N: ~58–112k daily bars across 111 tickers; ~28 names/quarter → tiny listwise queries; DL is overfit-prone by construction.
Cost calibration: half_spread_bps and eta are conservative floors, not central estimates; participation is a placeholder until real R$ volume lands.
Effective N is ~10× smaller than nominal due to overlapping 10d labels — handled by uniqueness weights and effective-N deflation, but it caps statistical power.

Reproduce

bash

uv sync
echo "BRAPI_TOKEN=your_key" > .env            # your own key; gitignored
uv run python -m stockprecog.brapi_ingest     # rebuild local panel (not shipped)
uv run python -m stockprecog.fundamental_tilt # champion (pre-registered)
uv run python -m stockprecog.rank_model       # LightGBM LambdaRank baseline
uv run python -m stockprecog.rank_tabpfn      # TabPFN v2 baseline (GPU)
uv run pytest -q                              # 26 tests (ruler invariants)

Citation

Paper (PDF): paper.pdf · DOI: 10.5281/zenodo.20706701 (concept DOI — latest version)

bibtex

@software{barbosa_stockprecog_2026,
  author    = {Barbosa, João Henrique},
  title     = {stockprecog: An honest-invalidation study of cross-sectional
               return prediction on the Brazilian equity market (B3)},
  year      = {2026},
  publisher = {Zenodo},
  version   = {v1.0.0},
  doi       = {10.5281/zenodo.20706701},
  url       = {https://doi.org/10.5281/zenodo.20706701}
}

JoaoHenriqueBarbosa/stockprecog

stockprecog — model card