褚宏光
0f8b9d836b
- Introduced a new "tilt" parameter to the strength scoring system, allowing for the assessment of triangle slope directionality. - Renamed existing parameters: "拟合贴合度" to "形态规则度" and "边界利用率" to "价格活跃度" for improved clarity. - Updated normalization methods for all strength components to ensure they remain within the [0, 1] range, facilitating LLM tuning. - Enhanced documentation to reflect changes in parameter names and scoring logic, including detailed explanations of the new tilt parameter. - Modified multiple source files and scripts to accommodate the new scoring structure and ensure backward compatibility. Files modified: - `src/converging_triangle.py`, `src/converging_triangle_optimized.py`, `src/triangle_detector_api.py`: Updated parameter names and scoring logic. - `scripts/plot_converging_triangles.py`, `scripts/generate_stock_viewer.py`: Adjusted for new scoring parameters in output. - New documentation files created to explain the renaming and new scoring system in detail.
138 lines
4.4 KiB
Markdown
138 lines
4.4 KiB
Markdown
# 强度分：形态规则度和价格活跃度的归一化详解
|
||
|
||
本文档详细说明**形态规则度** (`geometry_score`) 和**价格活跃度** (`activity_score`) 这两个强度分量的归一化实现方式。
|
||
|
||
---
|
||
|
||
## 4. 形态规则度 (`geometry_score`) 的归一化
|
||
|
||
**目的**：测量枢轴点（4-8个关键点）到拟合线的贴合程度，形态越规则得分越高。
|
||
|
||
### 计算步骤
|
||
|
||
```python
|
||
# 1. 计算相对误差：遍历所有枢轴点
|
||
for i in range(n):
|
||
fitted_value = slope * pivot_indices[i] + intercept
|
||
rel_error = abs(pivot_values[i] - fitted_value) / max(abs(fitted_value), 1e-9)
|
||
sum_rel_error += rel_error
|
||
|
||
# 2. 计算平均相对误差
|
||
mean_rel_error = sum_rel_error / n
|
||
|
||
# 3. 指数衰减归一化
|
||
SCALE_FACTOR = 20.0
|
||
geometry_score = np.exp(-mean_rel_error * SCALE_FACTOR)
|
||
|
||
# 4. 显式 clamp 到 [0, 1]
|
||
geometry_score = min(1.0, max(0.0, geometry_score))
|
||
```
|
||
|
||
### 归一化原理
|
||
|
||
- **输入**：`mean_rel_error` ∈ [0, +∞)（平均相对误差，0 = 完美拟合）
|
||
- **指数衰减**：`exp(-mean_rel_error * 20)` 将误差映射到 (0, 1] 区间
|
||
- 误差 = 0 → 得分 = 1.0（完美拟合）
|
||
- 误差 = 0.05 → 得分 ≈ 0.37（中等拟合）
|
||
- 误差 = 0.10 → 得分 ≈ 0.14（较差拟合）
|
||
- 误差 → ∞ → 得分 → 0（完全不拟合）
|
||
- **缩放因子 20**：决定衰减速度，值越大对误差越敏感
|
||
|
||
### 范围保证
|
||
|
||
- 指数函数 `exp(-x)` 对于 x ≥ 0，输出自然在 (0, 1] 区间
|
||
- 最后显式 clamp 确保异常情况下也在 [0, 1]
|
||
|
||
### 实现位置
|
||
|
||
- 函数名：`calc_geometry_score_numba()`
|
||
- 文件：`src/converging_triangle_optimized.py`（第350-374行）
|
||
|
||
---
|
||
|
||
## 5. 价格活跃度 (`activity_score`) 的归一化
|
||
|
||
**目的**：测量价格在通道内的振荡充分性，识别真实博弈 vs 僵尸形态。
|
||
|
||
### 计算步骤
|
||
|
||
```python
|
||
# 1. 逐日计算活跃度（遍历 240 天完整数据）
|
||
total_activity = 0.0
|
||
valid_days = 0
|
||
|
||
for i in range(start, end + 1):
|
||
# 1.1 计算当日通道宽度
|
||
upper_line = upper_slope * i + upper_intercept
|
||
lower_line = lower_slope * i + lower_intercept
|
||
channel_width = upper_line - lower_line
|
||
|
||
if channel_width <= 0:
|
||
continue
|
||
|
||
# 1.2 计算空白距离
|
||
dist_to_upper = max(0.0, upper_line - high[i]) # 高点未触及上沿的距离
|
||
dist_to_lower = max(0.0, low[i] - lower_line) # 低点未触及下沿的距离
|
||
|
||
# 1.3 计算空白比例
|
||
blank_ratio = (dist_to_upper + dist_to_lower) / channel_width
|
||
|
||
# 1.4 单日活跃度 = 1 - 空白比例，并 clamp 到 [0, 1]
|
||
day_activity = max(0.0, min(1.0, 1.0 - blank_ratio))
|
||
|
||
total_activity += day_activity
|
||
valid_days += 1
|
||
|
||
# 2. 计算平均活跃度（已自动在 [0, 1] 区间）
|
||
activity_score = total_activity / valid_days
|
||
```
|
||
|
||
### 归一化原理
|
||
|
||
- **输入**：每日的 `blank_ratio`（空白比例）∈ [0, +∞)
|
||
- 0 = 价格完全填满通道（高点触上沿且低点触下沿）
|
||
- 1 = 价格只占通道一半空间
|
||
- \> 1 = 价格严重偏离通道（理论上不应出现）
|
||
- **反转**：`1 - blank_ratio` 将"空白"转为"活跃"
|
||
- blank_ratio = 0 → activity = 1（最活跃）
|
||
- blank_ratio = 0.5 → activity = 0.5（中等活跃）
|
||
- blank_ratio = 1 → activity = 0（不活跃）
|
||
- **双重 clamp**：
|
||
1. **每日 clamp**：`max(0.0, min(1.0, ...))` 确保单日得分在 [0, 1]
|
||
2. **最终平均**：因为每日都在 [0, 1]，平均值自然在 [0, 1]
|
||
|
||
### 范围保证
|
||
|
||
- 每日活跃度通过双边 clamp 严格限制在 [0, 1]
|
||
- 最终得分是所有有效日的平均值，数学上保证在 [0, 1]
|
||
|
||
### 实际意义
|
||
|
||
- **0.8-1.0**：价格充分振荡，真实博弈形态
|
||
- **0.5-0.8**：价格较活跃，形态有效
|
||
- **0.2-0.5**：价格偏弱，形态存疑
|
||
- **< 0.2**：僵尸形态，触发空白惩罚机制
|
||
|
||
### 实现位置
|
||
|
||
- 函数名：`calc_activity_score_numba()`
|
||
- 文件：`src/converging_triangle_optimized.py`（第378-412行）
|
||
|
||
---
|
||
|
||
## 总结
|
||
|
||
这两个分量的归一化方式各有特点：
|
||
|
||
1. **形态规则度**：使用指数衰减映射，对误差敏感度高，适合质量评估
|
||
2. **价格活跃度**：使用线性反转+双重clamp，逐日计算后取平均，适合统计性指标
|
||
|
||
两者都严格保证输出在 [0, 1] 区间，满足强度分系统的设计要求。
|
||
|
||
---
|
||
|
||
**文档创建日期**：2026-01-29
|
||
**相关文档**：
|
||
- `docs/强度分组成梳理.md`：6个强度分量的完整说明
|
||
- `discuss/20260129-讨论.md`：强度分参数讨论
|