Préparer une stratégie de trading

Article Machine Learning for Trading

Résumé

Cette revue de chapitre décrit une démarche fondée sur des hypothèses pour définir des stratégies de trading avant le backtest. Elle relie les hypothèses documentées sur les données et une configuration immuable à l’analyse exploratoire, aux études d’événements et à une fiche structurée de stratégie. Cette fiche consigne une hypothèse réfutable, un plan de mise en œuvre, les exigences de faisabilité et un plan de validation, afin de transformer les observations de recherche en spécifications pouvant être examinées et versionnées.

Le contenu passe en revue des exemples de momentum sur ETF, de retour à la moyenne des taux de financement crypto, de microstructure intrajournalière et de données historiques sur les facteurs. Il souligne des pratiques respectant l’information disponible à chaque instant, telles que le décalage des données macroéconomiques, les statistiques glissantes, les jointures temporelles de rendements futurs et les règles de temporisation des études d’événements. Les exemples statistiques incluent des tests du rendement moyen ajustés selon Newey-West et des estimations d’incertitude du ratio de Sharpe. Il s’agit d’une revue de l’architecture et du code, et non d’un nouveau test empirique : les évaluations de qualité et recommandations relèvent du jugement de l’auteur, et les résultats exploratoires décrits n’établissent pas de rentabilité en réel ou hors échantillon.

Idées clés

Une stratégie doit partir d’une hypothèse réfutable et d’hypothèses documentées sur les données.
Une fiche structurée peut consigner les exigences de mise en œuvre, de faisabilité et de validation dans un format versionné.
Les jointures temporelles avec tolérance gèrent plus sûrement les observations irrégulières que des décalages fixes supposés.
Les décalages respectant l’information disponible à chaque instant et les calculs glissants contribuent à réduire le biais d’anticipation pendant l’exploration.
L’analyse exploratoire et les revues de qualité du code n’établissent pas la performance hors échantillon.

Étiquettes

Momentum Retour à la moyenne Statistiques Backtest Apprentissage automatique Stratégies événementielles

Texte intégral

# Chapter 5: Strategy Definition - Codebase Analysis


# Chapter 5: Strategy Definition - Codebase Analysis

**Generated**: 2026-01-25
**Analysis Type**: Deep Architecture & Pattern Review

---

## Executive Summary

Chapter 5 is a **mature, well-architected** educational codebase demonstrating strategy pre-specification. The code quality is high with consistent patterns, strong documentation, and proper separation of concerns. The architecture follows a **hypothesis-driven exploration** pattern rather than a production trading system pattern.

**Overall Quality**: ★★★★☆ (4/5) - Publication-ready with minor improvements possible

---

## Architecture Overview

```
06_strategy_definition/
├── chapter/                    # Book manuscript (section-based workflow)
│   ├── draft.md               # Auto-generated from sections
│   ├── summary.md             # Chapter scope definition
│   ├── bibliography.md        # Academic references
│   └── sections/              # 10 section files (ground truth)
│
├── code/                      # 5 Jupytext notebooks (~4,500 LOC)
│   ├── 01_etf_momentum.py     # ETF correlation, regime, rotation (858 lines)
│   ├── 02_crypto_premium.py   # Funding rate mean reversion (792 lines)
│   ├── 03_algoseek_intraday.py # Microstructure exploration (817 lines)
│   ├── 04_aqr_factor_performance.py # Century of factor evidence (1,395 lines)
│   └── 05_strategy_term_sheet_template.py # Python dataclasses (694 lines)
│
├── term_sheets/               # 8 completed strategy specifications
├── catalog/                   # 5 notebook metadata JSON files
├── figures/                   # Book figures (AI-generated + notebook)
└── reviews/                   # Code review artifacts
```

### Pattern: Hypothesis-Driven Exploration

The codebase follows a clear pedagogical pattern:

```
1. Data Contract (documented assumptions)
      ↓
2. Configuration Dataclass (frozen, versioned)
      ↓
3. Exploratory Analysis (correlation, distribution, regime)
      ↓
4. Event Study (with cooldown, forward returns)
      ↓
5. Key Priors Summary (for Term Sheet)
```

---

## Code Quality Assessment

### Strengths

| Aspect | Score | Evidence |
|--------|-------|----------|
| **Documentation** | ★★★★★ | Every notebook has purpose, data contract, assumptions, limitations |
| **Type Safety** | ★★★★☆ | Frozen dataclasses, type hints in key functions |
| **Polars Usage** | ★★★★★ | Proper lazy API, `.over()` windows, no pandas except boundaries |
| **Time Handling** | ★★★★★ | Time-based joins (not shifts), timezone-aware, tolerance parameters |
| **Point-in-Time** | ★★★★☆ | Rolling z-scores, macro lag handling, clear look-ahead warnings |

### Patterns Observed

#### 1. Configuration Dataclasses (Excellent)

```python
@dataclass(frozen=True)
class EtfExplorationConfig:
    """Configuration for ETF momentum exploration (Chapter 2)."""
    start_date_str: str = "2007-01-01"
    trading_days_per_year: int = 252
    momentum_formation_days: int = 126  # 6 months
    skip_month_days: int = 21  # Skip most recent month
    # ... etc
```

**Benefit**: Immutable, documented, single source of truth for notebook parameters.

#### 2. Time-Based Forward Returns (Production-Quality)

```python
def _compute_forward_returns_hourly(
    df: pl.DataFrame,
    *,
    price_col: str,
    horizons_hours: tuple[int, ...],
    group_col: str,
    ts_col: str,
    tolerance_hours: int = 2,
) -> pl.DataFrame:
```

**Pattern**: Uses `join_asof` with tolerance instead of `.shift()` to handle gaps correctly.

#### 3. Event Study with Proper Cooldown

```python
def event_study_with_cooldown(
    df: pl.DataFrame,
    regime_col: str = "regime",
    cooldown_hours: int = 24,
) -> pl.DataFrame:
```

**Pattern**: Stateful thinning that compares to last KEPT event, not previous event.

#### 4. Strategy Term Sheet as Python Dataclass

```python
@dataclass
class StrategyTermSheet:
    name: str
    version: str
    status: Literal["Draft", "Review", "Approved"]
    classification: Literal["Price-Based", "Fundamental", "Structural", "ML-Enhanced"]
    hypothesis: FalsifiableHypothesis
    blueprint: ImplementationBlueprint
    feasibility: FeasibilityGate
    validation: ValidationPlan
```

**Benefit**: Enforces structure, enables JSON/YAML export, supports version control.

---

## Component Inventory

### Notebooks (5)

| Notebook | Purpose | Data Sources | Lines | Quality |
|----------|---------|--------------|-------|---------|
| `01_etf_momentum` | Correlation, regime, rotation | ETF Universe, FRED | 858 | ★★★★★ |
| `02_crypto_premium` | Funding rate mean reversion | Binance Premium/OHLCV | 792 | ★★★★★ |
| `03_algoseek_intraday` | Microstructure exploration | AlgoSeek NASDAQ100 | 817 | ★★★★☆ |
| `04_aqr_factor_performance` | Century of factor evidence | AQR, Fama-French | 1,395 | ★★★★★ |
| `05_strategy_term_sheet_template` | Interactive Term Sheet | None | 694 | ★★★★★ |

### Term Sheets (8)

| Term Sheet | Classification | Status |
|------------|----------------|--------|
| `etf_momentum__rotational_momentum.md` | Price-Based | Complete |
| `crypto_premium__premium_funding_reversal.md` | Structural | Complete |
| `nasdaq100_reversal__intraday_orderflow_reversal.md` | Structural | Complete |
| `us_factors__cross_sectional_factor_momentum.md` | Price-Based | Complete |
| `futures_carry__term_structure_momentum.md` | Structural | Complete |
| `fx_momentum__cross_sectional_momentum.md` | Price-Based | Complete |
| `algoseek_sp500__options_volatility_alpha.md` | Structural | Complete |
| `etf_momentum.md` | Price-Based | Legacy (use etf_momentum version) |

### Chapter Sections (10)

```
00_preamble.md
01_the_strategy_specification_problem.md
02_a_map_of_strategies_and_edges.md
03_the_strategy_term_sheet.md
04_filling_the_term_sheet.md
05_minimum_specification_feasibility.md
06_validation_plan_and_failure_conditions.md
07_where_ideas_come_from.md
08_evidence_discipline_and_the_factor_zoo.md
09_key_takeaways.md
```

---

## Data Flow Architecture

```
                    ┌─────────────────┐
                    │   Data Layer    │
                    │ (utils, DATA_DIR)│
                    └────────┬────────┘
                             │
        ┌────────────────────┼────────────────────┐
        │                    │                    │
        ▼                    ▼                    ▼
┌───────────────┐  ┌───────────────┐  ┌───────────────┐
│  ETF Universe │  │ Crypto Premium│  │   AlgoSeek    │
│   + FRED      │  │   + OHLCV     │  │  Minute Bars  │
└───────┬───────┘  └───────┬───────┘  └───────┬───────┘
        │                  │                  │
        ▼                  ▼                  ▼
┌───────────────────────────────────────────────────┐
│              Exploration Notebooks                │
│  (correlation, regime, event study, statistics)   │
└───────────────────────┬───────────────────────────┘
                        │
                        ▼
┌───────────────────────────────────────────────────┐
│               Strategy Term Sheets                │
│    (hypothesis, blueprint, feasibility, validation)│
└───────────────────────────────────────────────────┘
                        │
                        ▼
               Downstream Chapters
           (Ch7 features, Ch17 backtest)
```

---

## Statistical Methods Used

### Factor Analysis (04_aqr_factor_performance.py)

| Method | Implementation | Quality |
|--------|----------------|---------|
| Newey-West HAC t-stat | `calculate_mean_return_tstat()` | ★★★★★ |
| Lo (2002) Sharpe SE | `calculate_sharpe_stats()` | ★★★★★ |
| Max Drawdown | `calculate_max_drawdown()` | ★★★★☆ |
| Rolling Correlation | `pl.rolling_corr()` native | ★★★★★ |

**Notable**: Proper distinction between Harvey et al. (2016) mean-return t-stat (for factor discovery) and Lo (2002) Sharpe ratio significance (for performance).

### Momentum Analysis (01_etf_momentum.py)

| Method | Implementation | Quality |
|--------|----------------|---------|
| Cross-sectional quintile ranking | `.rank("ordinal").over("timestamp")` | ★★★★★ |
| 6-1 Momentum (skip-month) | Vectorized with `.shift()` | ★★★★★ |
| Yield curve regime filter | FRED data with 2-bday lag | ★★★★★ |
| Rotation simulation | Month-by-month loop (pedagogical) | ★★★★☆ |

### Event Studies (02_crypto_premium.py, 03_algoseek_intraday.py)

| Method | Implementation | Quality |
|--------|----------------|---------|
| Rolling z-score | Point-in-time (168h window) | ★★★★★ |
| Event cooldown | Stateful per-group thinning | ★★★★★ |
| Forward returns | Time-based join with tolerance | ★★★★★ |
| Winsorization | Global quantiles (exploration only) | ★★★★☆ |

---

## Improvement Recommendations

### Priority 1: Minor Code Improvements

1. **Add `__all__` exports to notebooks** for potential module reuse:
   ```python
   __all__ = ["EtfExplorationConfig", "calculate_momentum_score"]
   ```

2. **Extract common utilities** to shared module:
   - `_compute_forward_returns_*` functions appear in multiple notebooks
   - `_event_cooldown_filter` is duplicated
   - Could live in `utils/exploration.py`

3. **Add catalog entry for `etf_momentum.md` legacy term sheet** or remove:
   - Currently `etf_momentum.md` exists alongside `etf_momentum__rotational_momentum.md`
   - Confusing which is canonical

### Priority 2: Documentation Enhancements

1. **Add notebook execution order** to README.md:
   ```markdown
   ## Execution Order
   1. 01_etf_momentum.py (ETF data exploration)
   2. 02_crypto_premium.py (crypto case study)
   ...
   ```

2. **Document term sheet naming convention**:
   - Pattern: `{dataset}__{strategy_type}.md`
   - Examples: `etf_momentum__rotational_momentum.md`, `crypto_premium__premium_funding_reversal.md`

### Priority 3: Testing Coverage

1. **Add unit tests for statistical functions**:
   - `calculate_mean_return_tstat()` - verify against known values
   - `calculate_sharpe_stats()` - test Lo (2002) SE calculation
   - `_event_cooldown_filter()` - edge cases

2. **Add data contract validation tests**:
   - Verify parquet schemas match documented contracts
   - Check for expected columns and dtypes

---

## Risk Areas

### Low Risk (Acceptable)

| Area | Observation | Mitigation |
|------|-------------|------------|
| Pandas boundary | Matplotlib/scipy require pandas | Minimal, conversion at boundary only |
| Global winsorization | Uses full-sample quantiles | Documented as exploration-only |
| Loop in rotation sim | Month-by-month for clarity | Acceptable for pedagogy |

### Medium Risk (Monitor)

| Area | Observation | Mitigation |
|------|-------------|------------|
| Duplicate term sheets | `etf_momentum.md` vs `etf_momentum__*.md` | Clarify canonical version |
| Hardcoded magic numbers | Some thresholds in notebook body | Move to config dataclass |
| Missing type hints | Helper functions lack full typing | Add for production reuse |

---

## Dependency Analysis

### External Dependencies

```python
# Core
polars          # DataFrames (primary)
numpy           # Numerical operations
pandas          # Boundary conversions only
plotly          # Visualization

# Domain
scipy.stats     # Statistical tests
IPython.display # Jupyter rendering

# ML4T
utils           # DATA_DIR, paths
ml4t.data.providers  # AQR, Fama-French
```

### Internal Dependencies

```
05_strategy_term_sheet_template.py
    ↓ exports
StrategyTermSheet (dataclass)
    ↓ used by
term_sheets/*.md (via to_markdown())
```

---

## Alignment with Book Standards

| Standard | Compliance | Notes |
|----------|------------|-------|
| Polars-first | ✅ Full | No pandas except visualization boundaries |
| Frozen config dataclass | ✅ Full | All notebooks use this pattern |
| Time-based joins | ✅ Full | No `.shift()` for forward returns |
| Data contracts documented | ✅ Full | Every notebook has contract table |
| Output to `figures/` only | ✅ Full | No ad-hoc file saves |
| No code in chapter text | ✅ Full | References notebooks by name |
| TEST mode support | ⚠️ Partial | Not all notebooks have explicit TEST guards |

---

## Summary

Chapter 5 demonstrates **exemplary educational code architecture**:

1. **Clear separation**: Data contracts → Config → Analysis → Summary
2. **Proper statistics**: Newey-West HAC, Lo Sharpe SE, time-based forward returns
3. **Point-in-time discipline**: Rolling z-scores, macro lag handling
4. **Versioned artifacts**: Term sheets as structured documents

**Recommended actions**:
- Extract common utilities to reduce duplication
- Clarify canonical term sheet naming
- Add unit tests for statistical functions

**Overall**: Ready for publication with minor cleanup.

---

*Analysis performed by Claude Code /development:analyze*

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.

Définir les stratégies de trading avant le backtest : recherche et validation

Résumé

Idées clés

Étiquettes