US : stratégie transversale sur actions, validation, coûts et échec hors échantillon

Article Machine Learning for Trading

Résumé

Cette étude évalue des signaux transversaux quotidiens sur un vaste univers d’actions US et décrit un long processus de recherche, des données disponibles à la date considérée et des variables construites jusqu’à la comparaison des modèles, la construction du portefeuille, les coûts et l’évaluation hors échantillon. Elle utilise une validation walk-forward, plusieurs horizons de rendements futurs, des modèles linéaires et à arbres, des modèles neuronaux tabulaires et temporels, ainsi que des méthodes à facteurs latents. La stratégie de trading classe les actions pour former des portefeuilles acheteur-vendeur neutres en dollars, en tenant compte des coûts de transaction variables selon les périodes et des coûts d’emprunt.

Les résultats de validation rapportés montrent des statistiques positives pour les signaux et la stratégie, avec des intervalles de confiance et des diagnostics ajustés pour la sélection. Cependant, la période hors échantillon 2016–2018 inverse la performance de la stratégie : son ratio de Sharpe est négatif, et les comparaisons appariées montrent une dégradation significative par rapport à la validation et à l’indice équipondéré. L’étude indique aussi que les coûts réduisent les rendements et que l’exécution constitue une contrainte opérationnelle. Ces résultats concernent un seul univers historique, une lignée de modèles, une fréquence de rééquilibrage et un régime hors échantillon ; la solidité des résultats de validation ne prouve pas une rentabilité durable, comme le montre la période hors échantillon elle-même.

Idées clés

L’étude vérifie si de faibles signaux au niveau des actions peuvent s’agréger sur un large univers transversal.
Le processus combine des variables financières et temporelles avec plusieurs familles de modèles et une validation walk-forward.
La stratégie forme des portefeuilles acheteur-vendeur neutres en dollars au quotidien et inclut des coûts de trading significatifs.
Les métriques de signal et de stratégie sont positives en validation et présentées avec des mesures d’incertitude et des ajustements de sélection.
La performance se dégrade significativement pendant la période hors échantillon 2016–2018, ce qui limite les affirmations d’un avantage persistant.

Étiquettes

Actions Apprentissage automatique Momentum Retour à la moyenne Backtest Exécution des ordres

Texte intégral

# Case Study: US Equities Panel


# Case Study: US Equities Panel

This case study is the broadest cross-sectional equity workflow in the book. It uses daily OHLCV data from NASDAQ Data Link for ~3,200 US stocks spanning 1990 through 2018-Q1 to test whether weak per-stock signals translate into a tradable strategy when scaled across thousands of names. The Fundamental Law of Active Management is the operating frame: the per-stock edge is small, but breadth across the cross-section is supposed to compensate. The role of this case study is to hold that claim against measured signal quality, paired-bootstrap confidence intervals, and an explicit holdout window.

The pipeline is unusually long because the universe is unusually large. Sixteen walk-forward folds (10y train, 1y validation), the most folds of any case study, are paired with multi-horizon labels and a feature panel that mixes momentum, mean-reversion, volatility, liquidity, value proxies, and walk-forward temporal models. The strategy is a daily long-short top-K cross-sectional ranker with dollar-neutral construction and material era-dependent costs (15-30 bps pre-decimalization, 5-15 bps after). The question the strategy-analysis notebook answers is whether the gross signal that survives this much testing also survives selection-adjusted resampling and the 2016-2018 holdout.

## At a Glance

| Property | Value |
|----------|-------|
| Asset Class | Broad US equities (NYSE/NASDAQ/AMEX) |
| Frequency | Daily |
| Universe | ~3,200 stocks (price > $5, ADV > $1M, point-in-time) |
| History | 1990 -- 2018-Q1 |
| Primary Label | fwd_ret_1d |
| CV Folds | 16 (10Y train, 1Y val) |
| Cost Model | Material (5-30 bps per leg, era-dependent + borrow) |

## Pipeline

| Stage | Notebook | Chapter | Description | Writes |
|-------|----------|---------|-------------|--------|
| Feasibility | [`01_feasibility_analysis`](01_feasibility_analysis.ipynb) | Ch6 | Universe breadth per decision date, cost regime, move-to-cost scale, walk-forward folds | Nothing - the evidence stays in the notebook |
| Labels | [`02_labels`](02_labels.ipynb) | Ch7 | 1-day, 5-day, and 21-day forward returns | `labels/fwd_ret_1d.parquet`, `labels/fwd_ret_5d.parquet`, `labels/fwd_ret_21d.parquet`, each with a `.digest.json` sidecar |
| Features | [`03_financial_features`](03_financial_features.ipynb) | Ch8 | 62 cross-sectional factors: momentum, volatility, liquidity, value | `features/financial.parquet` |
| Temporal | [`04_model_based_features`](04_model_based_features.ipynb) | Ch9 | Walk-forward Wasserstein regime distance, FFD, GARCH features | `features/model_based.parquet` |
| Evaluation | [`05_evaluation`](05_evaluation.ipynb) | Ch7--9 | Feature-label IC diagnostics across the full panel | `evaluation/triage_ledger.parquet`, `evaluation/ic_timeseries.parquet` |
| Linear | [`06_linear`](06_linear.ipynb) | Ch11 | Ridge, LASSO, ElasticNet on the full feature matrix | Training runs and prediction sets in `run_log/registry.db`; coefficients under `run_log/training/{hash}/`, scores under `run_log/predictions/{hash}/` |
| GBM | [`07_gbm`](07_gbm.ipynb) | Ch12 | LightGBM grid across leaf profiles and loss functions | Training runs and prediction sets; boosters, `learning_curves.parquet`, and `fold_metrics.parquet` under `run_log/training/{hash}/` |
| Tabular DL | [`08_tabular_dl`](08_tabular_dl.ipynb) | Ch12 | TabM attention-style ensembling on the cross-section | Training runs and prediction sets; checkpoints under `run_log/training/tabular_dl/` |
| NLinear | [`09_dl_nlinear`](09_dl_nlinear.ipynb) | Ch13 | Minimal temporal baseline with last-value normalization | Training runs and prediction sets; checkpoints under `run_log/training/deep_learning/` |
| LSTM | [`10_dl_lstm`](10_dl_lstm.ipynb) | Ch13 | Sequential memory across daily return windows | Training runs and prediction sets; checkpoints under `run_log/training/deep_learning/` |
| TSMixer | [`11_dl_tsmixer`](11_dl_tsmixer.ipynb) | Ch13 | Time-mixing and feature-mixing across the 60-day lookback | Training runs and prediction sets; checkpoints under `run_log/training/deep_learning/` |
| Weekly DL | [`12_dl_weekly`](12_dl_weekly.ipynb) | Ch13 | Weekly-cadence LSTM/NLinear comparison | Training runs and prediction sets; checkpoints under `run_log/training/deep_learning/` |
| Latent Factors | [`13_latent_factors`](13_latent_factors.ipynb) | Ch14 | Index notebook for PCA + IPCA on the broad equity panel | Nothing - it reads the registry |
| PCA | [`13a_pca`](13a_pca.ipynb) | Ch14 | Static factor extraction from the return covariance | Training runs and prediction sets |
| IPCA | [`13b_ipca`](13b_ipca.ipynb) | Ch14 | Instrumented PCA with characteristic-conditioned loadings | Training runs and prediction sets |
| Causal DML | [`14_causal_dml`](14_causal_dml.ipynb) | Ch15 | Causal effect of 12-1 momentum on daily returns | A row in the registry's `causal_runs` |
| Model Analysis | [`15_model_analysis`](15_model_analysis.ipynb) | -- | Cross-model IC comparison and fold stability diagnostics | Nothing - it reads the registry |
| Backtest | [`16_backtest`](16_backtest.ipynb) | Ch16 | Daily long-short top-K strategy simulation | One backtest run per prediction set and entry scheme; `daily_returns.parquet`, `weights.parquet`, `trades.parquet`, `fills.parquet`, `equity.parquet`, `portfolio_state.parquet`, and `spec.json` under `run_log/backtest/{hash}/` |
| Portfolio | [`17_portfolio_management`](17_portfolio_management.ipynb) | Ch17 | Allocation sweep on the highest-IC GBM signal | One backtest run per allocation method, same artifact layout |
| Risk | [`18_risk_management`](18_risk_management.ipynb) | Ch19 | Position-level and portfolio-level risk overlays | One backtest run per overlay variant, same artifact layout |
| Costs | [`19_costs`](19_costs.ipynb) | Ch18 | Cost-grid sweep on the strategies the three earlier stages produced | One backtest run per cost level, same artifact layout |
| Holdout Predictions | [`20_holdout_predictions`](20_holdout_predictions.ipynb) | Ch20 | Refit of the selected configuration on history ending before the holdout window | One training run and one prediction set at `split='holdout'` |
| Holdout Backtest | [`21_holdout_backtest`](21_holdout_backtest.ipynb) | Ch20 | The holdout predictions traded under the selected allocator, overlay and cost level | One backtest run at `stage='holdout'`, same artifact layout |
| Strategy Analysis | [`22_strategy_analysis`](22_strategy_analysis.ipynb) | Ch20 | End-to-end strategy assessment: signal, lineage, holdout, attribution | `results/strategy_assessment.json`, `20_strategy_synthesis/output/us_equities_panel/us_equities_panel_tearsheet.html` |

## Key Results

**Signal direction.** GBM `leaves_31_huber` on the 5-day variant horizon achieves the highest cross-stage validation Sharpe and a strong daily-pooled IC on the panel's fwd_ret_1d grid. Pooled IC is 0.0357 with the HAC-adjusted 95% CI at [0.0293, 0.0421] over 4,018 daily cross-sections (t_HAC = 10.96), well clear of zero. Per-family rank-1 IC is monotone in horizon for GBM (1d 0.032 → 5d 0.043 → 21d 0.058) and linear (1d 0.016 → 5d 0.022 → 21d 0.029), with each CI excluding zero. Tree-based and linear families produce signals with low pairwise correlation, so an ensemble across families would not be fighting a single shared signal.

**Strategy-stage performance with CIs.** Validation Sharpe for this lineage's risk_overlay carrier (score_weighted top_k=20 + `time_exit_40`) is 2.028 with a paired-bootstrap 95% CI of [1.464, 2.549] (PSR p ≈ 2e-15, classification `excludes_zero_strong`). The strategy posts a higher Sharpe than the equal-weight US-equities universe over the same window by 1.11 [0.48, 1.76] (p ≈ 0, `excludes_zero_strong`). A FF5+MOM HAC regression credits the validation edge as alpha-driven: annualized alpha ≈ 0.76 with t_HAC ≈ 7.7, residual Sharpe ≈ 2.04, R² ≈ 0.01. Cohort-level selection-bias metrics from `cohort_metrics` (family cohort `risk_overlay/fwd_ret_5d/gbm`, K_variants = 20 position-level overlays, K_eff_MP ≈ 2.0, K_eff_ER ≈ 2.5) record DSR_ER 0.106 with p ≈ 0 (and DSR_MP 0.112, ER and MP within 0.006), and PBO 0.0 across 12,870 CSCV combinations × 16 folds. On the broader label cohort (`label/fwd_ret_5d`, K_variants = 314 across all families and stages) the same leader records DSR_ER 0.065 at p ≈ 7e-13 with K_eff_ER ≈ 12.2 — the leader's edge survives both the small overlay-cohort adjustment and the cross-stage cross-family adjustment.

**Holdout closure.** The 2016-Q1 to 2018-Q1 holdout puts this lineage at Sharpe −0.492. The index-paired diff against validation reads −2.520 [−3.804, −1.117] with p ≈ 5e-4. The CI excludes zero on the negative side, so the deterioration is statistically resolved. The reference equal-weight universe over the same holdout window posts Sharpe 1.71 (well above its validation reading of 0.92, reflecting the cap-weighted bull market of that period); against that elevated reference, strategy minus benchmark over the holdout reads −2.363 [−4.591, −0.213] (p ≈ 0.03, `excludes_zero_strong` on the negative side). The overall reading is that the validation edge does not carry across the holdout regime under the chosen rebalance cadence.

**Friction floor.** The cost-sensitivity sweep on this lineage produces a moderately steep envelope. Within the cross-stage rank-1 prediction lineage, zero-cost gross Sharpe is 2.503 (the gross-return ceiling under the 5-day-label / daily-marked strategy), and Sharpe at the 10 bps post-decimalization midpoint is 2.117. The edge-to-cost ratio comfortably clears the 1.2× kill-condition floor (`evidence_passes`). The universal Ch20 gates resolve: validation Sharpe lower bound (1.46) is above zero, and the holdout strategy-vs-EW CI excludes zero negatively. The steepness of the cost curve and the daily rebalance cadence place the strategy in a regime where execution quality is the binding operational constraint.

## Running

```bash
# From repo root
uv run python case_studies/us_equities_panel/01_feasibility_analysis.py
uv run python case_studies/us_equities_panel/02_labels.py
uv run python case_studies/us_equities_panel/03_financial_features.py
uv run python case_studies/us_equities_panel/04_model_based_features.py
uv run python case_studies/us_equities_panel/05_evaluation.py
uv run python case_studies/us_equities_panel/06_linear.py
uv run python case_studies/us_equities_panel/07_gbm.py
uv run python case_studies/us_equities_panel/08_tabular_dl.py
uv run python case_studies/us_equities_panel/09_dl_nlinear.py
uv run python case_studies/us_equities_panel/10_dl_lstm.py
uv run python case_studies/us_equities_panel/11_dl_tsmixer.py
uv run python case_studies/us_equities_panel/12_dl_weekly.py
uv run python case_studies/us_equities_panel/13_latent_factors.py
uv run python case_studies/us_equities_panel/13a_pca.py
uv run python case_studies/us_equities_panel/13b_ipca.py
uv run python case_studies/us_equities_panel/14_causal_dml.py
uv run python case_studies/us_equities_panel/15_model_analysis.py
uv run python case_studies/us_equities_panel/16_backtest.py
uv run python case_studies/us_equities_panel/17_portfolio_management.py
uv run python case_studies/us_equities_panel/18_risk_management.py
uv run python case_studies/us_equities_panel/19_costs.py
uv run python case_studies/us_equities_panel/20_holdout_predictions.py
uv run python case_studies/us_equities_panel/21_holdout_backtest.py
uv run python case_studies/us_equities_panel/22_strategy_analysis.py
```

The strategy-analysis notebook in `22_strategy_analysis.py` writes a full diagnostic tear sheet (`template="full"`) to the case study's gitignored output directory; readers regenerate it locally.

## Run Log

Model training runs, predictions, and backtest results are tracked in a content-addressed registry under `run_log/registry.db`.

Reproduit dans son intégralité avec attribution, conformément à la licence de la source. Licence: MIT

Ce résumé a été rédigé par l’agent de recherche de Stratmill à partir de la source originale ; il n’en est pas une copie.