debajyotidasgupta/repro-label-horizon-paradox-artifacts

Reproduction: The Label Horizon Paradox

Independent reproduction of ICML 2026 paper #9730 — The Label Horizon Paradox: Rethinking Supervision Targets in Financial Forecasting (Chen-Hui Song, Shuoling Liu, Liyuan Chen; E Fund Management). arXiv 2602.03395 · OpenReview G43CIfmmxh.

Part of the Hugging Face × AlphaXiv ICML-2026 reproduction challenge.

Why this is a mechanism reproduction

The paper's benchmark uses proprietary CSI 300/500/1000 minute-level data (E Fund Management) and ships no code, so the exact Table-1 numbers cannot be reproduced by anyone outside the lab. However, the paper's Appendix C derives a complete, self-contained generative model (a time-varying APT process) from which the paradox, its theory, and the bi-level method all follow. We reproduce those core scientific claims by sampling from that model — this is a faithful test of the paper's mechanism, not the closed benchmark.

What reproduces (and what does not)

Claim	Result
C1 three regimes (decline / near-equality / hump)	✅ reproduced; empirical OLS vs closed-form theory corr > 0.94
C2 Theorem 3.2 + Corollary 3.3 factorization	✅ reproduced on CSI 300/500/1000 proxies; corr 0.96–0.99
C3 bi-level learns per-horizon λ → δ*	✅ reproduced; learned λ recovers all three regimes
C4 Table 1 universal improvement (10 backbones)	✅ direction reproduced (Ours ≥ Standard); exact CSI numbers ❌ infeasible (proprietary data)
C5 data spans + 15-min patch construction	✅ spec verified + pipeline reproduces 16×D (A-share) / 26×D (US) tensors

Layout

src/lhp/            reproduction package
  market_sim.py     time-varying APT generative model (Appendix C) + closed-form theory
  models.py         the 10 backbones (LSTM,GRU,DLinear,RLinear,PatchTST,iTransformer,Mamba,BiMamba+,ModernTCN,TCN)
  metrics.py        IC / ICIR / RankIC / RankICIR / TopRet / Sharpe (Appendix B.2)
  trainer.py        standard training + bi-level horizon learning (Section 4)
experiments/        claim{1..5}_*.py + run_all.py driver
outputs/            result JSON / CSV / interactive HTML figures per claim
make_poster.py      builds the reproduction poster
build_logbook.py    populates the Trackio logbook claim pages
poster.html/png     rendered poster

Reproduce

pip install torch numpy scipy scikit-learn plotly pandas matplotlib
python experiments/run_all.py --scale smoke --device cpu     # seconds, sanity
python experiments/run_all.py --scale full  --device cpu     # full local reproduction

Device notes. The bi-level meta-gradient needs double-backward. Use CPU (works for all backbones) — Apple MPS lacks RNN double-backward, and CUDA's cuDNN RNN also lacks it (the native fallback is numerically unreliable for second-order gradients). The non-bi-level claims (C1, C2, C5) run fine on GPU/MPS.

Compute

HF Jobs was unavailable for this account (402 Payment Required, no credits). Claims 1, 2, 5 ran on a Vast.ai RTX 2080 Ti (CUDA); the bi-level claims 3, 4 ran on CPU for the double-backward reason above. Total GPU spend ≈ $0.68.

Faithfulness notes

• Bi-level hyperparameters follow the paper: 3-epoch warm-up on the standardized mean-field label, inner fine-tune lr 1e-6, outer lr 1e-3, entropy γ=1e-3, softmax λ, 20-day support/query batches.
• Claim 3 uses the RLinear backbone (one of the paper's ten). The paper's Fig. 4 uses LSTM; the bi-level mechanism is backbone-agnostic, and RLinear gives clean, reproducible concentration across all three regimes in the reduced-data synthetic setting.
• The Fig-4 warm-up ablation (collapse to earliest horizons without warm-up) depends on real-market feature/label temporal proximity absent from the whitened-factor proxy, so it is not reproduced here.