feat(weight): 实现 Kelly 仓位权重模式
- config_loader.py: WeightType 枚举新增 KELLY - simple_rotation.py: compute_position_weights 新增 kelly 分支 - 公式: w_i = max(score_i, 0) / sum(max(score_j, 0)) - 负分自动排除 (Kelly: 不下注负期望) - 全负分时 fallback 到等权 - _generate_signals 传递 scores 给 kelly 模式 - config_simple.yaml: weight 改为 kelly - 新增策略总结文档: kelly_weight.md 回测对比 (2020-2026): - equal: 年化 19.88%, 夏普 1.13, 回撤 -14.65% - rank: 年化 22.90%, 夏普 1.12, 回撤 -16.27% - kelly: 年化 30.13%, 夏普 1.15, 回撤 -20.44%
This commit is contained in:
@@ -0,0 +1,213 @@
|
||||
# Kelly 仓位权重模式实现总结
|
||||
|
||||
**日期**: 2026-06-07
|
||||
**Commit**: `06d12f2`
|
||||
**相关文件**:
|
||||
- `rotation/simple_rotation.py`
|
||||
- `rotation/config_loader.py`
|
||||
- `rotation/config_simple.yaml`
|
||||
|
||||
---
|
||||
|
||||
## 1. 背景与动机
|
||||
|
||||
### 1.1 问题提出
|
||||
原有仓位管理支持两种模式：
|
||||
- **equal**: 等权分配 (1/N)
|
||||
- **rank**: 按排名三角权重 (第1名50%, 第2名33%, 第3名17%)
|
||||
|
||||
用户询问能否使用 Kelly 准则进行仓位分配。
|
||||
|
||||
### 1.2 Kelly 准则简介
|
||||
经典 Kelly 公式: **f* = W - (1-W)/R**
|
||||
- W = 胜率（历史盈利交易占比）
|
||||
- R = 盈亏比（平均盈利/平均亏损）
|
||||
|
||||
### 1.3 经典 Kelly 的挑战
|
||||
|
||||
| 问题 | 说明 |
|
||||
|------|------|
|
||||
| 样本量不足 | 每个标的被持有的天数有限，统计胜率/盈亏比不稳定 |
|
||||
| 非平稳性 | 市场环境变化导致历史统计不代表未来 |
|
||||
| 极端值敏感 | 一次大亏会剧烈改变 Kelly 比例 |
|
||||
| 需要 expanding window | 回测中每天用截止当天的历史来估计，计算量大 |
|
||||
|
||||
---
|
||||
|
||||
## 2. 解决方案：Score-Proportional Kelly 近似
|
||||
|
||||
### 2.1 核心思路
|
||||
利用当前动量分数 `weighted_momentum_score = annualized_return × R²` 作为 edge 代理，构造 Kelly 近似：
|
||||
|
||||
```
|
||||
w_i = max(score_i, 0) / Σ max(score_j, 0)
|
||||
```
|
||||
|
||||
### 2.2 设计优势
|
||||
- **无需额外历史统计**：每天从截面数据直接计算
|
||||
- **天然支持 expanding window**：每天用最新数据
|
||||
- **负分自动排除**：Kelly 原则 - 不下注负期望
|
||||
- **可插拔设计**：与现有 equal/rank 模式统一接口
|
||||
|
||||
### 2.3 公式推导
|
||||
动量分数 `score = annualized_return × R²` 包含：
|
||||
- **annualized_return**: 趋势方向和强度
|
||||
- **R²**: 趋势质量（信噪比）
|
||||
|
||||
正 score 意味着正期望，Kelly 建议按 edge 比例下注。归一化后得到仓位权重。
|
||||
|
||||
---
|
||||
|
||||
## 3. 代码实现
|
||||
|
||||
### 3.1 枚举扩展 (config_loader.py)
|
||||
```python
|
||||
class WeightType(str, Enum):
|
||||
"""仓位加权模式"""
|
||||
EQUAL = "equal" # 等权
|
||||
RANK = "rank" # 按排名加权
|
||||
KELLY = "kelly" # Kelly准则近似
|
||||
```
|
||||
|
||||
### 3.2 核心函数 (simple_rotation.py)
|
||||
```python
|
||||
def compute_position_weights(
|
||||
ranked_holdings: List[str],
|
||||
weight_type: str = 'equal',
|
||||
scores: Dict[str, float] = None, # 新增参数
|
||||
) -> Dict[str, float]:
|
||||
"""
|
||||
Schemes:
|
||||
equal: each slot = 1/N, duplicates summed.
|
||||
rank: slot i (0-indexed) = (N-i) / triangular(N), duplicates summed.
|
||||
kelly: w_i = max(score_i, 0) / sum(max(score_j, 0)).
|
||||
Score-proportional weighting as Kelly criterion proxy.
|
||||
Negative scores excluded (Kelly: don't bet on negative edge).
|
||||
"""
|
||||
N = len(ranked_holdings)
|
||||
if N == 0:
|
||||
return {}
|
||||
|
||||
weights: Dict[str, float] = {}
|
||||
|
||||
if weight_type == 'kelly':
|
||||
if not scores:
|
||||
raise ValueError("Kelly weighting requires 'scores' parameter")
|
||||
# Kelly proxy: weight proportional to positive scores
|
||||
positive_scores = {c: max(scores.get(c, 0.0), 0.0) for c in set(ranked_holdings)}
|
||||
total = sum(positive_scores.values())
|
||||
if total <= 0:
|
||||
# Fallback to equal if all scores non-positive
|
||||
w = 1.0 / len(positive_scores)
|
||||
for code in positive_scores:
|
||||
weights[code] = w
|
||||
else:
|
||||
for code in ranked_holdings:
|
||||
w = positive_scores.get(code, 0.0) / total
|
||||
weights[code] = weights.get(code, 0.0) + w
|
||||
|
||||
elif weight_type == 'rank':
|
||||
triangular = N * (N + 1) / 2
|
||||
for i, code in enumerate(ranked_holdings):
|
||||
w = (N - i) / triangular
|
||||
weights[code] = weights.get(code, 0.0) + w
|
||||
else:
|
||||
# equal (default)
|
||||
w = 1.0 / N
|
||||
for code in ranked_holdings:
|
||||
weights[code] = weights.get(code, 0.0) + w
|
||||
|
||||
return weights
|
||||
```
|
||||
|
||||
### 3.3 调用点修改
|
||||
```python
|
||||
# _generate_signals 中传递 scores
|
||||
self._pending_weights = compute_position_weights(
|
||||
ranked_holdings, self.weight_type, scores=factors,
|
||||
)
|
||||
```
|
||||
|
||||
### 3.4 配置使用 (config_simple.yaml)
|
||||
```yaml
|
||||
rotation:
|
||||
diversified: true
|
||||
select_num: 3
|
||||
weight: kelly # 可选: equal, rank, kelly
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 4. 回测结果对比
|
||||
|
||||
**回测区间**: 2020-01-10 ~ 2026-06-08 (1550 交易日)
|
||||
|
||||
| 指标 | equal | rank | **kelly** |
|
||||
|------|-------|------|-----------|
|
||||
| 累计收益 | 204.97% | 255.45% | **405.23%** |
|
||||
| 年化收益 | 19.88% | 22.90% | **30.13%** |
|
||||
| 最大回撤 | **-14.65%** | -16.27% | -20.44% |
|
||||
| 夏普比率 | 1.13 | 1.12 | **1.15** |
|
||||
| Calmar比率 | 1.36 | 1.41 | **1.47** |
|
||||
| 日胜率 | 54.07% | 53.75% | **54.10%** |
|
||||
| 调仓次数 | 392 | 392 | 392 |
|
||||
|
||||
### 4.1 结果分析
|
||||
|
||||
**Kelly 模式特点**:
|
||||
- **收益最高**: 按动量分数比例分配权重，强势标的获得更大仓位
|
||||
- **夏普最高**: 风险调整后收益最优
|
||||
- **Calmar 最高**: 收益/回撤比最优
|
||||
- **回撤较大**: 集中度更高导致波动更大
|
||||
|
||||
**三种模式定位**:
|
||||
- **equal**: 保守型，分散风险，适合风险厌恶
|
||||
- **rank**: 平衡型，按排名阶梯分配
|
||||
- **kelly**: 进攻型，按 edge 比例集中配置
|
||||
|
||||
### 4.2 为什么日胜率会变化？
|
||||
|
||||
虽然信号生成（调仓日期、持仓标的）完全相同，但仓位权重影响每日组合收益：
|
||||
|
||||
```
|
||||
daily_return = Σ (weight_i × return_i)
|
||||
```
|
||||
|
||||
当某天收益接近 0 时，权重分配的变化可能让它在正/负之间翻转。例如：
|
||||
- 排名第1的标的大跌
|
||||
- rank 模式给 50% 权重 → 组合收益可能变负
|
||||
- equal 模式只给 33% → 影响较小，可能仍为正
|
||||
|
||||
---
|
||||
|
||||
## 5. 设计原则
|
||||
|
||||
### 5.1 可插拔架构
|
||||
- 统一函数签名，通过 `weight_type` 参数切换
|
||||
- 新增模式只需添加分支，不影响现有逻辑
|
||||
- `scores` 参数可选，仅 kelly 模式需要
|
||||
|
||||
### 5.2 防御性设计
|
||||
- Kelly 模式校验 `scores` 参数
|
||||
- 全负分时自动 fallback 到等权
|
||||
- 与 bond fill 机制兼容（债券 score 通常为负）
|
||||
|
||||
### 5.3 配置驱动
|
||||
- 通过 YAML 配置切换，无需修改代码
|
||||
- 支持环境变量覆盖
|
||||
- 与现有配置体系一致
|
||||
|
||||
---
|
||||
|
||||
## 6. 后续优化方向
|
||||
|
||||
1. **Half-Kelly**: 使用 f*/2 降低波动
|
||||
2. **动态 Kelly**: 根据市场状态调整 Kelly 系数
|
||||
3. **风险预算**: 结合波动率进行风险平价分配
|
||||
4. **多因子 Kelly**: 综合多个因子 score 计算 edge
|
||||
|
||||
---
|
||||
|
||||
## 7. 结论
|
||||
|
||||
Kelly 仓位模式通过 score-proportional 近似，在保持可插拔架构的同时，实现了最优风险调整后收益。对于追求收益最大化的场景，kelly 模式是首选；对于风险厌恶场景，equal 模式更稳健。
|
||||
Reference in New Issue
Block a user