stock-analysis-plugin 0.2.5 → 0.2.7
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +126 -15
- package/package.json +2 -1
- package/pi/index.ts +918 -72
- package/schemas/market_review_schema.json +12 -5
- package/schemas/report_schema.json +68 -4
- package/scripts/cli-timeout.mjs +27 -0
- package/scripts/json-input.mjs +26 -0
- package/scripts/setup-python.mjs +6 -6
- package/skills/bull-trend/SKILL.md +34 -10
- package/skills/bull-trend/references/fat-tails.md +91 -0
- package/skills/bull-trend/scripts/tail_analysis.py +132 -0
- package/skills/event-driven/SKILL.md +6 -0
- package/skills/growth-quality/SKILL.md +6 -0
- package/skills/market-review/SKILL.md +2 -0
- package/skills/stock-analysis/SKILL.md +25 -5
- package/skills/stock-analysis/references/stateless-analysis.md +66 -0
- package/skills/stock-screener/SKILL.md +7 -1
- package/skills/strategy-backtest/SKILL.md +13 -0
- package/skills/wisburg-research/SKILL.md +2 -0
- package/templates/market_review/full.md.j2 +7 -0
- package/templates/stock_analysis/_evidence.md.j2 +11 -0
- package/templates/stock_analysis/brief.md.j2 +6 -1
- package/templates/stock_analysis/full.md.j2 +8 -3
- package/tools/_subproc.py +66 -11
- package/tools/alert_rules.py +32 -4
- package/tools/backtest.py +95 -18
- package/tools/capabilities.py +11 -0
- package/tools/cli_budgets.json +26 -0
- package/tools/gather.py +24 -16
- package/tools/lixinger.py +100 -0
- package/tools/market_regime.py +7 -3
- package/tools/market_review.py +37 -15
- package/tools/position_context.py +20 -2
- package/tools/report_renderer.py +2 -0
- package/tools/risk_screening.py +45 -17
- package/tools/screener.py +74 -24
- package/tools/sec_edgar.py +346 -0
- package/tools/sec_holders.py +263 -0
- package/tools/signal_tracker.py +218 -24
- package/tools/stateless_analysis.py +630 -0
- package/tools/stock_data.py +170 -32
- package/tools/technical.py +16 -2
- package/tools/value_data.py +159 -0
- package/tools/watchlist.py +24 -1
- package/tools/watchlist_context.py +25 -10
|
@@ -18,17 +18,24 @@
|
|
|
18
18
|
"type": "object",
|
|
19
19
|
"description": "市场温度(来自 market_review.review)",
|
|
20
20
|
"properties": {
|
|
21
|
-
"score": { "type": "number", "minimum": 0, "maximum": 100 },
|
|
22
|
-
"level": { "type": "string", "enum": ["constructive", "neutral", "weak"] },
|
|
23
|
-
"signal": { "type": "string", "enum": ["green", "yellow", "red"] }
|
|
21
|
+
"score": { "type": ["number", "null"], "minimum": 0, "maximum": 100 },
|
|
22
|
+
"level": { "type": "string", "enum": ["constructive", "neutral", "weak", "unknown"] },
|
|
23
|
+
"signal": { "type": "string", "enum": ["green", "yellow", "red", "unknown"] }
|
|
24
24
|
},
|
|
25
25
|
"required": ["score", "signal"]
|
|
26
26
|
},
|
|
27
27
|
"strategy_stance": {
|
|
28
|
-
"type": "string",
|
|
29
|
-
"enum": ["offensive", "balanced", "defensive"],
|
|
28
|
+
"type": ["string", "null"],
|
|
29
|
+
"enum": ["offensive", "balanced", "defensive", null],
|
|
30
30
|
"description": "建议策略姿态"
|
|
31
31
|
},
|
|
32
|
+
"data_quality": {
|
|
33
|
+
"type": "object",
|
|
34
|
+
"properties": {
|
|
35
|
+
"flags": { "type": "array", "items": { "type": "string" } },
|
|
36
|
+
"limitations": { "type": "array", "items": { "type": "string" } }
|
|
37
|
+
}
|
|
38
|
+
},
|
|
32
39
|
"regime": {
|
|
33
40
|
"type": "object",
|
|
34
41
|
"description": "市场阶段(来自 detect_market_regime)",
|
|
@@ -48,9 +48,9 @@
|
|
|
48
48
|
"type": "object",
|
|
49
49
|
"description": "风险筛查结果(来自 screen_risk)",
|
|
50
50
|
"properties": {
|
|
51
|
-
"risk_level": { "type": "string", "enum": ["low", "medium", "high"] },
|
|
52
|
-
"risk_score": { "type": "integer", "minimum": 0, "maximum": 100 },
|
|
53
|
-
"veto_buy": { "type": "boolean" },
|
|
51
|
+
"risk_level": { "type": "string", "enum": ["low", "medium", "high", "unknown"] },
|
|
52
|
+
"risk_score": { "type": ["integer", "null"], "minimum": 0, "maximum": 100 },
|
|
53
|
+
"veto_buy": { "type": ["boolean", "null"] },
|
|
54
54
|
"flags": {
|
|
55
55
|
"type": "array",
|
|
56
56
|
"items": {
|
|
@@ -124,7 +124,71 @@
|
|
|
124
124
|
"action_checklist": { "type": "array", "items": { "type": "string" } }
|
|
125
125
|
}
|
|
126
126
|
},
|
|
127
|
-
"risk_warning": { "type": "array", "items": { "type": "string" } }
|
|
127
|
+
"risk_warning": { "type": "array", "items": { "type": "string" } },
|
|
128
|
+
"data_quality": {
|
|
129
|
+
"type": "object",
|
|
130
|
+
"properties": {
|
|
131
|
+
"flags": {
|
|
132
|
+
"type": "array",
|
|
133
|
+
"items": {
|
|
134
|
+
"type": "string",
|
|
135
|
+
"enum": [
|
|
136
|
+
"partial_data",
|
|
137
|
+
"stale_quote",
|
|
138
|
+
"conflicting_data",
|
|
139
|
+
"insufficient_sample"
|
|
140
|
+
]
|
|
141
|
+
}
|
|
142
|
+
},
|
|
143
|
+
"limitations": {
|
|
144
|
+
"type": "array",
|
|
145
|
+
"items": {
|
|
146
|
+
"type": "string"
|
|
147
|
+
}
|
|
148
|
+
}
|
|
149
|
+
},
|
|
150
|
+
"required": [
|
|
151
|
+
"flags",
|
|
152
|
+
"limitations"
|
|
153
|
+
]
|
|
154
|
+
},
|
|
155
|
+
"evidence": {
|
|
156
|
+
"type": "array",
|
|
157
|
+
"items": {
|
|
158
|
+
"type": "object",
|
|
159
|
+
"properties": {
|
|
160
|
+
"path": {
|
|
161
|
+
"type": "string",
|
|
162
|
+
"description": "原始工具结果路径,例如 get_quote.price;已知返回对象中未提供的末级字段只能以 value=null 表示未知,其 source/as_of 可为 null,若填写必须来自父对象;不得跨越缺失对象或数组编造路径。"
|
|
163
|
+
},
|
|
164
|
+
"label": {
|
|
165
|
+
"type": "string",
|
|
166
|
+
"description": "读者可理解的指标名"
|
|
167
|
+
},
|
|
168
|
+
"value": {},
|
|
169
|
+
"source": {
|
|
170
|
+
"description": "引用已返回字段时,沿 path 继承最近父对象的 source(该键缺失时用 snapshot_source);子对象来源覆盖祖先,显式 null 保持未知。数组项和警告字符串同样继承,不能仅因叶子没有 source 就填 null;整条路径未提供来源时才填 null。",
|
|
171
|
+
"type": [
|
|
172
|
+
"string",
|
|
173
|
+
"null"
|
|
174
|
+
]
|
|
175
|
+
},
|
|
176
|
+
"as_of": {
|
|
177
|
+
"description": "引用已返回字段时,沿 path 继承最近父对象的 data_as_of,子对象值(包括显式 null)覆盖祖先;数组项同样继承。整条路径均缺失时为 null,保留完整时间与时区。report_date 是报告期,不能代替此时间。",
|
|
178
|
+
"type": [
|
|
179
|
+
"string",
|
|
180
|
+
"null"
|
|
181
|
+
]
|
|
182
|
+
}
|
|
183
|
+
},
|
|
184
|
+
"required": [
|
|
185
|
+
"path",
|
|
186
|
+
"value",
|
|
187
|
+
"source",
|
|
188
|
+
"as_of"
|
|
189
|
+
]
|
|
190
|
+
}
|
|
191
|
+
}
|
|
128
192
|
},
|
|
129
193
|
"required": ["stock_name", "stock_code", "analysis_date", "sentiment_score",
|
|
130
194
|
"decision_type", "confidence", "core_conclusion"]
|
|
@@ -0,0 +1,27 @@
|
|
|
1
|
+
import budgets from "../tools/cli_budgets.json" with { type: "json" };
|
|
2
|
+
import { basename } from "node:path";
|
|
3
|
+
|
|
4
|
+
// Hosts leave the CLI time to return its own structured failure.
|
|
5
|
+
export function commandTimeout(script, args) {
|
|
6
|
+
const key = `${basename(script)}:${args[0] ?? ""}`;
|
|
7
|
+
let budget = budgets[key] ?? budgets.default;
|
|
8
|
+
const option = (name, fallback) => args.includes(name) ? args[args.indexOf(name) + 1] ?? fallback : fallback;
|
|
9
|
+
const positiveInt = (value, fallback) => /^\d+$/.test(String(value)) && Number(value) > 0 ? Number(value) : fallback;
|
|
10
|
+
if (key === "screener.py:screen") {
|
|
11
|
+
budget += Math.ceil(positiveInt(option("--top", 20), 20) / budgets.l2_workers) * budgets.l2_batch;
|
|
12
|
+
} else if (key === "watchlist.py:analyze" || key === "watchlist_context.py:build") {
|
|
13
|
+
const count = String(args[1] ?? "").split(",").filter(s => s.trim()).length;
|
|
14
|
+
const workers = positiveInt(option("--workers", 3), 3);
|
|
15
|
+
budget += Math.ceil(count / workers) * budgets.watchlist_item;
|
|
16
|
+
if (key === "watchlist_context.py:build") {
|
|
17
|
+
budget += Math.ceil(count / Math.min(workers * 2, 6)) * budgets["anomaly_detect.py:detect"];
|
|
18
|
+
if (args.includes("--include-market-review")) budget += 3 * budgets["market_review.py:review"] + 30;
|
|
19
|
+
}
|
|
20
|
+
} else if (key === "market_review.py:review" && option("--market", "A") === "all") {
|
|
21
|
+
budget = 3 * budget + 30;
|
|
22
|
+
} else if (key === "technical.py:analyze" && args.includes("--periods")) {
|
|
23
|
+
const periods = String(option("--periods", "daily")).split(",").map(p => p.trim()).filter(Boolean);
|
|
24
|
+
budget *= new Set([option("--period", "daily"), ...periods]).size;
|
|
25
|
+
}
|
|
26
|
+
return budget * 1000 + 30_000;
|
|
27
|
+
}
|
|
@@ -0,0 +1,26 @@
|
|
|
1
|
+
import { mkdtemp, writeFile, rm } from "node:fs/promises";
|
|
2
|
+
import { tmpdir } from "node:os";
|
|
3
|
+
import { join } from "node:path";
|
|
4
|
+
|
|
5
|
+
// Large reports/K-lines exceed Windows argv limits. Remove transport files even on failure.
|
|
6
|
+
export async function withJsonInput(input, execute) {
|
|
7
|
+
let directory;
|
|
8
|
+
let result;
|
|
9
|
+
try {
|
|
10
|
+
directory = await mkdtemp(join(tmpdir(), "stock-analysis-"));
|
|
11
|
+
const path = join(directory, "input.json");
|
|
12
|
+
await writeFile(path, JSON.stringify(input), { encoding: "utf8", mode: 0o600 });
|
|
13
|
+
result = await execute(path);
|
|
14
|
+
} catch {
|
|
15
|
+
result = JSON.stringify({ error: "Analysis command failed; check input and interpreter/dependencies" });
|
|
16
|
+
} finally {
|
|
17
|
+
if (directory) {
|
|
18
|
+
try {
|
|
19
|
+
await rm(directory, { recursive: true, force: true });
|
|
20
|
+
} catch {
|
|
21
|
+
result = JSON.stringify({ error: "Could not remove temporary analysis input" });
|
|
22
|
+
}
|
|
23
|
+
}
|
|
24
|
+
}
|
|
25
|
+
return result;
|
|
26
|
+
}
|
package/scripts/setup-python.mjs
CHANGED
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
#!/usr/bin/env node
|
|
2
|
-
import {
|
|
2
|
+
import { execFileSync } from "node:child_process";
|
|
3
3
|
import { existsSync } from "node:fs";
|
|
4
4
|
import { join } from "node:path";
|
|
5
5
|
import { fileURLToPath } from "node:url";
|
|
@@ -14,14 +14,14 @@ const venvDir = join(pkgDir, ".venv");
|
|
|
14
14
|
const venvPython = venvPythonPath(pkgDir);
|
|
15
15
|
const requirements = join(pkgDir, "tools", "requirements.txt");
|
|
16
16
|
|
|
17
|
-
function run(
|
|
18
|
-
|
|
17
|
+
function run(bin, args) {
|
|
18
|
+
execFileSync(bin, args, { stdio: "inherit", timeout: 1_800_000 });
|
|
19
19
|
}
|
|
20
20
|
|
|
21
21
|
function findPython() {
|
|
22
22
|
for (const name of ["python3", "python"]) {
|
|
23
23
|
try {
|
|
24
|
-
const ver =
|
|
24
|
+
const ver = execFileSync(name, ["--version"], { stdio: "pipe" }).toString().trim();
|
|
25
25
|
const match = ver.match(/(\d+)\.(\d+)/);
|
|
26
26
|
if (match && (parseInt(match[1]) > 3 || (parseInt(match[1]) === 3 && parseInt(match[2]) >= 10))) {
|
|
27
27
|
return name;
|
|
@@ -42,10 +42,10 @@ try {
|
|
|
42
42
|
}
|
|
43
43
|
|
|
44
44
|
if (!existsSync(venvDir)) {
|
|
45
|
-
run(
|
|
45
|
+
run(python, ["-m", "venv", venvDir]);
|
|
46
46
|
}
|
|
47
47
|
|
|
48
|
-
run(
|
|
48
|
+
run(venvPython, ["-m", "pip", "install", "-q", "-r", requirements]);
|
|
49
49
|
} catch (err) {
|
|
50
50
|
console.warn(
|
|
51
51
|
`\n[pi-stock-analysis] Python setup failed: ${err.message}\n` +
|
|
@@ -1,24 +1,30 @@
|
|
|
1
1
|
---
|
|
2
2
|
name: bull-trend
|
|
3
|
-
description:
|
|
3
|
+
description: 趋势交易:识别中长期趋势,制定入场、持仓和退出规则。用于趋势股持有、规则化趋势策略或肥尾收益分析。
|
|
4
4
|
---
|
|
5
5
|
|
|
6
6
|
# 趋势追踪分析
|
|
7
7
|
|
|
8
8
|
## 核心理念
|
|
9
9
|
|
|
10
|
-
|
|
10
|
+
趋势追踪用可执行的规则参与趋势,通过控制试错损失、保留盈利空间捕获少数大行情。肥尾描述收益分布,趋势描述价格路径;肥尾指标不是趋势启动或延续信号。
|
|
11
|
+
|
|
12
|
+
用户询问塔勒布、肥尾计算、极端行情或策略是否依赖少数大盈利时,先读取 [肥尾度量与趋势收益诊断](references/fat-tails.md)。概念问题直接解释方法;需要具体统计结果时,按参考中的数据口径准备输入并调用 `scripts/tail_analysis.py`。普通趋势判断无需加载这份参考。
|
|
13
|
+
|
|
14
|
+
仅解释肥尾概念时,到此完成;下述流程用于具体标的的趋势分析。
|
|
11
15
|
|
|
12
16
|
## 执行流程
|
|
13
17
|
|
|
14
18
|
### 1. 采集数据
|
|
15
19
|
|
|
16
|
-
运行数据采集脚本获取 120 根日K
|
|
20
|
+
运行数据采集脚本获取 120 根日K线及技术指标(相对本 skill 目录执行):
|
|
17
21
|
|
|
18
22
|
```bash
|
|
19
23
|
python3 scripts/gather.py <symbol>
|
|
20
24
|
```
|
|
21
25
|
|
|
26
|
+
确认数据日期、周期和复权口径;120 根日K线用于近期趋势判断。肥尾分支按参考文件另取历史收益,报告实际取得的区间和样本量。
|
|
27
|
+
|
|
22
28
|
### 2. 趋势确认
|
|
23
29
|
|
|
24
30
|
**多头趋势确认条件**:
|
|
@@ -32,26 +38,44 @@ python3 scripts/gather.py <symbol>
|
|
|
32
38
|
- MACD 在零轴之下运行
|
|
33
39
|
- 价格沿下降通道运行
|
|
34
40
|
|
|
41
|
+
报告哪些条件已满足、哪些尚未满足,并给出对应数值或价格位置。空头趋势对仅能做多的股票账户意味着退出或观望;做空建议须有对应市场和账户能力。
|
|
42
|
+
|
|
35
43
|
### 3. 趋势阶段判断
|
|
36
44
|
|
|
37
45
|
| 阶段 | 特征 | 操作 |
|
|
38
46
|
|------|------|------|
|
|
39
|
-
| 启动期 | 均线粘合后向上发散,放量突破 |
|
|
40
|
-
| 加速期 | 价格加速上涨,远离均线 |
|
|
41
|
-
| 成熟期 | 涨幅放缓,量能递减 |
|
|
42
|
-
| 衰退期 |
|
|
47
|
+
| 启动期 | 均线粘合后向上发散,放量突破 | 检查入场触发和初始风险 |
|
|
48
|
+
| 加速期 | 价格加速上涨,远离均线 | 按持仓规则跟踪,更新保护位 |
|
|
49
|
+
| 成熟期 | 涨幅放缓,量能递减 | 检查退出条件,阶段标签本身不触发减仓 |
|
|
50
|
+
| 衰退期 | 跌破预先选定的均线或趋势支撑 | 按退出规则执行 |
|
|
43
51
|
|
|
44
52
|
### 4. 关键指标
|
|
45
53
|
|
|
46
54
|
- **ADX 替代**:用 MA 斜率和排列程度衡量趋势强度
|
|
47
55
|
- **趋势线**:连接近期低点作为上升趋势线
|
|
48
|
-
-
|
|
56
|
+
- **回调幅度**:描述相对前一波涨幅的回撤及是否破坏支撑;斐波那契比例只作为观察位置,不能单独证明回调健康。
|
|
57
|
+
|
|
58
|
+
### 5. 形成交易规则
|
|
59
|
+
|
|
60
|
+
先读取 [短线交易决策框架](../stock-analysis/references/short-term-trading.md)的风险单位和市场成交约束,再形成同一套入场、持仓和退出规则:
|
|
61
|
+
|
|
62
|
+
- **入场**:指定周期、触发条件、判断时点和成交假设。例如收盘突破此前 20 根K线最高价,阈值排除当前K线;历史检验使用信号产生后可成交的价格。
|
|
63
|
+
- **初始风险**:给出初始止损、单笔风险预算和仓位。按预算 / 入场到止损的每股损失估算数量,并受资金、交易单位和组合风险限制;跳空和无法成交可能使实际亏损超过预算。
|
|
64
|
+
- **持仓**:趋势条件持续满足时保留盈利空间。保护位随有利走势更新,做多止损只上移;涨幅大或收益处于高分位本身不构成退出条件。
|
|
65
|
+
- **退出**:明确采用均线失效、通道破位或移动止损中的哪一种、如何触发。固定止盈属于可选策略约束,要说明它可能截断大盈利;参数是待检验假设,不归为塔勒布公式。
|
|
66
|
+
|
|
67
|
+
规则完成标准:每项都有可核验的触发条件,说明下一次观察时点,以及触发后动作。用户已有规则时按其规则评价,调整另列为建议。
|
|
68
|
+
|
|
69
|
+
### 6. 按需检验
|
|
70
|
+
|
|
71
|
+
用户要求历史验证时使用 [策略回测](../strategy-backtest/SKILL.md)。先核对现有 DSL 能否表达规则,无法表达的条件标为未验证;不以近似规则冒充原策略。趋势收益的利润集中度诊断按肥尾参考文件执行。
|
|
49
72
|
|
|
50
73
|
## 输出要求
|
|
51
74
|
|
|
52
75
|
分析完成后输出:
|
|
53
76
|
1. 当前趋势方向和阶段
|
|
54
|
-
2.
|
|
77
|
+
2. 趋势证据和未满足条件
|
|
55
78
|
3. 关键趋势支撑/压力位
|
|
56
79
|
4. 入场/持有/离场建议
|
|
57
|
-
5.
|
|
80
|
+
5. 初始风险、保护位和退出触发条件
|
|
81
|
+
6. 数据或成交限制;肥尾分支附计算口径和不确定性
|
|
@@ -0,0 +1,91 @@
|
|
|
1
|
+
# 肥尾度量与趋势收益诊断
|
|
2
|
+
|
|
3
|
+
本参考供趋势 skill 和趋势策略回测按需读取。数值由本 skill 的 [计算脚本](../scripts/tail_analysis.py) 执行,无需独立工具注册。保留输入、运行参数和输出;没有执行环境或数据时只解释方法,数值标为未计算。
|
|
4
|
+
|
|
5
|
+
## 调用脚本
|
|
6
|
+
|
|
7
|
+
从插件根目录使用安装了现有 numpy 依赖的 Python 运行(POSIX 通常为 `.venv/bin/python3`,Windows 为 `.venv/Scripts/python.exe`):
|
|
8
|
+
|
|
9
|
+
```bash
|
|
10
|
+
python skills/bull-trend/scripts/tail_analysis.py returns historical-returns.json --n 2 10 30 --draws 20000 --seed 42
|
|
11
|
+
python skills/bull-trend/scripts/tail_analysis.py returns historical-returns.json --current 0.05
|
|
12
|
+
python skills/bull-trend/scripts/tail_analysis.py trades backtest-result.json
|
|
13
|
+
```
|
|
14
|
+
|
|
15
|
+
`returns` 输入是同周期历史对数收益的 JSON 数组,如 `[0.01, -0.02, 0.005]`,以小数表示;`--current` 的当前收益另传,必须从历史输入中排除。脚本不转换价格、不抓行情、不识别日期,因此先按下文校验并记录数据口径。`trades` 输入是含完整 `trades` 数组的回测 JSON。脚本 stdout 输出 JSON,失败返回 `error` 并以非零状态退出。
|
|
16
|
+
|
|
17
|
+
κ 默认 n0 = 1,以固定种子做 iid 经验重抽样,输出每个 n 的 κ 和平均绝对偏差;换种子重复运行检查 Monte Carlo 波动,换历史窗口检查数据敏感性。输出没有总体置信区间。`sample_count` 是历史收益数量,`draws` 是模拟轮数,两者不能混称样本量。`current_event.percentile` 为 0..1,`upper_exceedances` / `lower_exceedances` 是历史严格超过 / 低于当前收益的次数。
|
|
18
|
+
|
|
19
|
+
## 先确定分析对象
|
|
20
|
+
|
|
21
|
+
| 用户问题 | 分析对象 | 可交付结论 |
|
|
22
|
+
|---|---|---|
|
|
23
|
+
| 收益分布有多肥尾 | 固定周期的历史收益 | 样本尾部特征、κ 的条件性估计 |
|
|
24
|
+
| 当前涨跌是否极端 | 当前收益与此前同周期收益 | 历史分位或阈值超越,不能推导后续方向 |
|
|
25
|
+
| 趋势系统是否靠大赢家 | 策略平仓盈亏 | 盈利集中度及移除大赢家后的敏感性 |
|
|
26
|
+
|
|
27
|
+
市场收益的厚尾、策略收益的正偏和趋势延续是不同性质。塔勒布的度量回答统计稳定性问题;下文的分位和盈利贡献是本项目的诊断方法,不称为塔勒布交易信号。
|
|
28
|
+
|
|
29
|
+
## 数据口径
|
|
30
|
+
|
|
31
|
+
使用 `get_kline` 或现有 `stock_data.py kline`,核对实际返回的日期范围、周期、数量和复权标记。股票历史价格保持一致复权口径,指数保持原始点位;先处理乱序、重复日期和无效价格,说明停牌或缺失导致的时间间隔,保留真实极端收益。
|
|
32
|
+
|
|
33
|
+
默认分析对数收益 `r_t = log(P_t / P_{t-1})`,仅使用正价格。日、周收益分别计算;周收益从周末价格或同一周日对数收益之和取得,避免把重叠多日收益当作独立样本。策略盈利贡献使用货币盈亏,不混用价格收益。
|
|
34
|
+
|
|
35
|
+
申请多年的同口径历史,报告实际样本量;120 根日K线只能给出短窗口描述。不存在通用的“够用样本数”;长历史也可能混合不同制度或波动阶段,按分段结果检查稳定性。
|
|
36
|
+
|
|
37
|
+
## 当前事件的历史分位
|
|
38
|
+
|
|
39
|
+
以当前事件之前的同周期收益 `r_1,...,r_N` 为参考,对当前收益 `r_*` 计算经验分位:
|
|
40
|
+
|
|
41
|
+
```text
|
|
42
|
+
F_hat(r_*) = count(r_i <= r_*) / N
|
|
43
|
+
```
|
|
44
|
+
|
|
45
|
+
报告单侧方向、参考区间、N 和超越次数。例如当前上涨高于历史样本的 99% 分位,仅说明它相对该样本少见。样本之外的事件频率不可由零超越次数认定为零;回测每个时点只使用此前数据计算阈值。
|
|
46
|
+
|
|
47
|
+
## 塔勒布的 κ:有限样本聚合速度
|
|
48
|
+
|
|
49
|
+
来源:[Taleb, How Much Data Do You Need? A Pre-asymptotic Metric for Fat-tailedness](https://arxiv.org/html/1802.05495),Definition 1、Equation (1) 和 III-B。
|
|
50
|
+
|
|
51
|
+
设 `X_i` 独立同分布且具有有限绝对一阶矩:
|
|
52
|
+
|
|
53
|
+
```text
|
|
54
|
+
S_n = sum(X_i, i=1..n)
|
|
55
|
+
M(n) = E[abs(S_n - E[S_n])]
|
|
56
|
+
kappa(n0, n) = 2 - log(n/n0) / log(M(n)/M(n0)), n > n0 >= 1
|
|
57
|
+
```
|
|
58
|
+
|
|
59
|
+
`M(n)` 是 n 项之和相对其均值的平均绝对偏差,不是中位数绝对偏差,也不是 n 项均值的偏差。高斯分布满足 `M(n)/M(n0) = sqrt(n/n0)`,因此 κ = 0。在论文适用的分布中,较高 κ 表示增加样本带来的均值稳定性改善较慢;κ 与 n0、n 有关。
|
|
60
|
+
|
|
61
|
+
脚本计算的是**基于历史经验分布的条件性估计**:用历史均值中心化后计算 `M_hat(1)`,有放回独立抽样 n 项相加,以绝对和值的模拟均值估计 `M_hat(n)`,代入 n0 = 1 的公式。
|
|
62
|
+
|
|
63
|
+
零离散度或零对数分母时,脚本返回 `kappa: null` 和 `unavailable_reason`。估计超出预期范围时报告原值并检查假设和数值稳定性,不强行裁剪到 [0,1]。
|
|
64
|
+
|
|
65
|
+
股价收益可能具有自相关、波动聚集和制度变化,独立抽样会抹去时间结构。按真实连续时间聚合或使用块抽样时,应将结果标为包含时间依赖的描述性统计,与原论文的 iid κ 区分。仅重抽历史不能创造尚未出现的极端事件,也不能证明总体尾部已被充分观测。
|
|
66
|
+
|
|
67
|
+
κ 不能区分获利方向,不能预测趋势启动。论文中 κ > 0.15 是对正态近似可靠性的启发式提醒,不是入场、止盈或“已到肥尾”的通用阈值。
|
|
68
|
+
|
|
69
|
+
## 趋势策略的盈利集中度
|
|
70
|
+
|
|
71
|
+
来源:[CFM, Making fat right tails fatter with trend following… most of the time](https://www.cfm.com/making-fat-right-tails-fatter-with-trend-following-most-of-the-time/)。趋势追踪的凸性和正偏依赖策略与时间尺度;单次跳空或快速反转未必能被系统捕获。
|
|
72
|
+
|
|
73
|
+
使用 `run_backtest` 返回的 `trades`,仅取 `type == "sell"` 的平仓记录,以 `pnl` 的货币金额计算;`pnl_pct` 是小数比例,不能当百分数或与货币盈亏相加。先确认返回结果是否完整、是否存在未平仓头寸以及费用覆盖口径。
|
|
74
|
+
|
|
75
|
+
```text
|
|
76
|
+
G = sum(max(pnl_i, 0))
|
|
77
|
+
L = sum(min(pnl_i, 0))
|
|
78
|
+
T = G + L
|
|
79
|
+
C_k = 最大 k 笔正盈利之和 / G
|
|
80
|
+
T_without_k = T - 最大 k 笔正盈利之和
|
|
81
|
+
```
|
|
82
|
+
|
|
83
|
+
默认报告最大 1 笔和最大 3 笔,正盈利笔数不足时使用实际数量并注明。G = 0 时集中度不可计算;同时报告交易笔数、G、L、最差单笔及回测已有的最大回撤。使用 G 作分母,避免总净利润接近零时比率失真。
|
|
84
|
+
|
|
85
|
+
`T_without_k` 是已实现盈亏的删除敏感性分析,保留其余交易金额不变;它不是删除交易后重新模拟的资金曲线或收益率。未平仓盈亏另列,不能把 T 当作账户总收益。少数赢家贡献高符合某些趋势系统的结构,但不能单凭这一点证明策略有效或失效;结合成本、回撤、不同区间和样本外表现评价。
|
|
86
|
+
|
|
87
|
+
现有回测有固定止盈默认值;省略 `take_profit` 不会关闭止盈。验证保留右尾的退出规则前,检查实际引擎和 DSL 的支持情况,将无法表达的规则明确列为未验证。
|
|
88
|
+
|
|
89
|
+
## 完成标准
|
|
90
|
+
|
|
91
|
+
报告指出分析对象、数据来源及区间、周期、复权口径、样本量、公式与估计方法、实际数值或未计算原因。κ 结果还需注明 n0、n、抽样假设和敏感性;盈利集中度需注明平仓与未平仓口径。交易建议依据趋势和风险规则,统计结果仅提供其能支持的证据。
|
|
@@ -0,0 +1,132 @@
|
|
|
1
|
+
#!/usr/bin/env python3
|
|
2
|
+
"""Skill-local tail diagnostics from historical returns or backtest JSON."""
|
|
3
|
+
|
|
4
|
+
import argparse
|
|
5
|
+
import json
|
|
6
|
+
import math
|
|
7
|
+
import sys
|
|
8
|
+
from pathlib import Path
|
|
9
|
+
|
|
10
|
+
import numpy as np
|
|
11
|
+
|
|
12
|
+
|
|
13
|
+
class Parser(argparse.ArgumentParser):
|
|
14
|
+
def error(self, message):
|
|
15
|
+
raise ValueError(message)
|
|
16
|
+
|
|
17
|
+
|
|
18
|
+
def finite_number(value):
|
|
19
|
+
if isinstance(value, bool) or not isinstance(value, (int, float)) or not math.isfinite(value):
|
|
20
|
+
raise ValueError("Input values must be finite JSON numbers")
|
|
21
|
+
return float(value)
|
|
22
|
+
|
|
23
|
+
|
|
24
|
+
def analyze_returns(data, args):
|
|
25
|
+
if not isinstance(data, list) or len(data) < 2:
|
|
26
|
+
raise ValueError("Expected a JSON array of at least two historical returns")
|
|
27
|
+
values = np.array([finite_number(value) for value in data])
|
|
28
|
+
if not 100 <= args.draws <= 1000000 or args.seed < 0 or any(n < 2 or n > 10000 for n in args.n):
|
|
29
|
+
raise ValueError("Require draws in 100..1000000, seed >= 0, and n in 2..10000")
|
|
30
|
+
centered = values - values.mean()
|
|
31
|
+
mad = float(np.abs(centered).mean())
|
|
32
|
+
rng = np.random.default_rng(args.seed)
|
|
33
|
+
rows = []
|
|
34
|
+
for n in args.n:
|
|
35
|
+
total = 0.0
|
|
36
|
+
if mad > 0:
|
|
37
|
+
# Bound temporary arrays when either summands or draws is large.
|
|
38
|
+
batch_size = max(1, 1000000 // n)
|
|
39
|
+
for start in range(0, args.draws, batch_size):
|
|
40
|
+
batch = min(batch_size, args.draws - start)
|
|
41
|
+
sums = rng.choice(centered, size=(batch, n)).sum(axis=1)
|
|
42
|
+
total += float(np.abs(sums).sum())
|
|
43
|
+
summed_mad = total / args.draws
|
|
44
|
+
denominator = math.log(summed_mad / mad) if mad > 0 and summed_mad > 0 else 0
|
|
45
|
+
rows.append(
|
|
46
|
+
{
|
|
47
|
+
"n0": 1,
|
|
48
|
+
"n": n,
|
|
49
|
+
"mean_absolute_deviation_sum": summed_mad,
|
|
50
|
+
"kappa": 2 - math.log(n) / denominator if denominator else None,
|
|
51
|
+
"unavailable_reason": None if denominator else "Zero dispersion or degenerate aggregation",
|
|
52
|
+
}
|
|
53
|
+
)
|
|
54
|
+
event = None
|
|
55
|
+
if args.current is not None:
|
|
56
|
+
current = finite_number(args.current)
|
|
57
|
+
event = {
|
|
58
|
+
"return": current,
|
|
59
|
+
"percentile": float(np.mean(values <= current)),
|
|
60
|
+
"upper_exceedances": int(np.sum(values > current)),
|
|
61
|
+
"lower_exceedances": int(np.sum(values < current)),
|
|
62
|
+
}
|
|
63
|
+
return {
|
|
64
|
+
"sample_count": len(values),
|
|
65
|
+
"draws": args.draws,
|
|
66
|
+
"seed": args.seed,
|
|
67
|
+
"mean_absolute_deviation": mad,
|
|
68
|
+
"kappa": rows,
|
|
69
|
+
"current_event": event,
|
|
70
|
+
"method": "iid resampling of centered empirical returns; not a population tail estimate or trading signal",
|
|
71
|
+
}
|
|
72
|
+
|
|
73
|
+
|
|
74
|
+
def analyze_trades(data):
|
|
75
|
+
if not isinstance(data, dict) or not isinstance(data.get("trades"), list):
|
|
76
|
+
raise ValueError("Expected a backtest JSON object with a trades array")
|
|
77
|
+
if any(not isinstance(trade, dict) or trade.get("type") not in {"buy", "sell"} for trade in data["trades"]):
|
|
78
|
+
raise ValueError("Each trade must have type buy or sell")
|
|
79
|
+
pnl = [finite_number(trade.get("pnl")) for trade in data["trades"] if trade["type"] == "sell"]
|
|
80
|
+
winners = sorted((value for value in pnl if value > 0), reverse=True)
|
|
81
|
+
gross_profit = math.fsum(winners)
|
|
82
|
+
gross_loss = math.fsum(value for value in pnl if value < 0)
|
|
83
|
+
net = math.fsum(pnl)
|
|
84
|
+
return {
|
|
85
|
+
"closed_trades": len(pnl),
|
|
86
|
+
"winning_trades": len(winners),
|
|
87
|
+
"gross_profit": gross_profit,
|
|
88
|
+
"gross_loss": gross_loss,
|
|
89
|
+
"realized_net_pnl": net,
|
|
90
|
+
"worst_trade_pnl": min(pnl) if pnl else None,
|
|
91
|
+
"winner_contribution": [
|
|
92
|
+
{
|
|
93
|
+
"requested_k": k,
|
|
94
|
+
"actual_k": min(k, len(winners)),
|
|
95
|
+
"share_of_gross_profit": math.fsum(winners[:k]) / gross_profit if gross_profit else None,
|
|
96
|
+
"realized_pnl_without_top_k": net - math.fsum(winners[:k]),
|
|
97
|
+
}
|
|
98
|
+
for k in (1, 3)
|
|
99
|
+
],
|
|
100
|
+
"method": "closed monetary pnl only; deletion sensitivity is not an equity-curve rerun",
|
|
101
|
+
}
|
|
102
|
+
|
|
103
|
+
|
|
104
|
+
def main():
|
|
105
|
+
try:
|
|
106
|
+
parser = Parser(description=__doc__)
|
|
107
|
+
sub = parser.add_subparsers(dest="command", required=True, parser_class=Parser)
|
|
108
|
+
returns = sub.add_parser("returns", help="Analyze a JSON array of historical log returns")
|
|
109
|
+
returns.add_argument("input", type=Path)
|
|
110
|
+
returns.add_argument("--n", nargs="+", type=int, default=[2, 10, 30])
|
|
111
|
+
returns.add_argument("--draws", type=int, default=20000)
|
|
112
|
+
returns.add_argument("--seed", type=int, default=42)
|
|
113
|
+
returns.add_argument("--current", type=float, help="Separate current return, excluded from historical input")
|
|
114
|
+
trades = sub.add_parser("trades", help="Analyze a complete run_backtest JSON result")
|
|
115
|
+
trades.add_argument("input", type=Path)
|
|
116
|
+
args = parser.parse_args()
|
|
117
|
+
data = json.loads(args.input.read_text(encoding="utf-8"))
|
|
118
|
+
result = analyze_returns(data, args) if args.command == "returns" else analyze_trades(data)
|
|
119
|
+
print(json.dumps(result, ensure_ascii=True, allow_nan=False))
|
|
120
|
+
return 0
|
|
121
|
+
except OSError:
|
|
122
|
+
error = "Cannot read input JSON file"
|
|
123
|
+
except (ValueError, OverflowError) as exc:
|
|
124
|
+
error = "Invalid JSON input" if isinstance(exc, json.JSONDecodeError) else str(exc)
|
|
125
|
+
except Exception:
|
|
126
|
+
error = "Tail analysis failed unexpectedly"
|
|
127
|
+
print(json.dumps({"error": error}, ensure_ascii=True))
|
|
128
|
+
return 1
|
|
129
|
+
|
|
130
|
+
|
|
131
|
+
if __name__ == "__main__":
|
|
132
|
+
sys.exit(main())
|
|
@@ -5,6 +5,10 @@ description: 事件驱动分析:评估公告、业绩、政策、订单或资
|
|
|
5
5
|
|
|
6
6
|
# 事件驱动分析
|
|
7
7
|
|
|
8
|
+
## 无状态分析扩展
|
|
9
|
+
|
|
10
|
+
历史事件窗口验证或公告原文变化分析时,MUST 读取[无状态研究工具](../stock-analysis/references/stateless-analysis.md),使用 analyze_event_study/analyze_disclosure_changes。事件日期、盘前盘后、价格窗口和正文覆盖必须核实。
|
|
11
|
+
|
|
8
12
|
## 输入数据说明
|
|
9
13
|
|
|
10
14
|
先运行:
|
|
@@ -22,6 +26,8 @@ gather 工具返回:
|
|
|
22
26
|
- `stock_info` — 公司基本信息
|
|
23
27
|
- `sector_rankings` — 板块排名
|
|
24
28
|
|
|
29
|
+
A 股事件原文先调用 `get_announcements`:业绩预告用 `forecast`、快报用 `express`、减持用 `reduction`、增发用 `issuance`、问询及回复用 `inquiry`、分红用 `dividend`,指定 `start_date/end_date`。返回 PDF 链接而非正文;读取原文后再量化,`truncated=true` 时缩小日期范围。
|
|
30
|
+
|
|
25
31
|
新闻摘要只用于发现事件。金额、期限、合同对手、业绩变化或政策原文必须由公告/原文确认;只有摘要时不做精确量化。
|
|
26
32
|
|
|
27
33
|
制定短线计划时读取[短线交易决策框架](../stock-analysis/references/short-term-trading.md),用状态机、相对强弱和风险单位约束事件交易。
|
|
@@ -5,6 +5,10 @@ description: 成长质量分析:用多期收入、利润、现金流和盈利
|
|
|
5
5
|
|
|
6
6
|
# 成长质量分析
|
|
7
7
|
|
|
8
|
+
## 无状态分析扩展
|
|
9
|
+
|
|
10
|
+
财报前后期变化或同行验证时,MUST 读取[无状态研究工具](../stock-analysis/references/stateless-analysis.md)中的同行和公告财报比较,使用 compare_stocks/analyze_disclosure_changes 整理同口径指标与正文证据,再解释经济含义。
|
|
11
|
+
|
|
8
12
|
## 输入数据说明
|
|
9
13
|
|
|
10
14
|
先运行:
|
|
@@ -22,6 +26,8 @@ gather 工具返回:
|
|
|
22
26
|
- `stock_info` — 公司基本信息(行业、板块)
|
|
23
27
|
- `sector_rankings` — 板块排名
|
|
24
28
|
|
|
29
|
+
需要财务趋势时调用 `get_financials(symbol, annual=true, periods=10)`。港股配置 `LIXINGER_TOKEN` 后优先理杏仁非金融公司财报,按真实 `report_type` 判断年报,保留实际财年末、披露日期、币种;历史 ROE/毛利率/净利率为百分数,顶层 `roe/profit_margin` 延续小数比例。读取 `available_periods/warnings`,缺期不补造。有效 Yahoo 顶层快照与历史行分开读:`currency` 为快照财报币种、`market_cap_currency` 为市值币种,历史币种在 `periods` 每行,`snapshot_source` 标记快照来源;null 不用历史值补成跨期或跨币种数字;顶层 `report_date` 未知时读取 `periods` 的历史日期,不当作当前快照报告期。
|
|
30
|
+
|
|
25
31
|
`financials` 通常只是单期摘要。多期趋势、经营现金流、应收账款或同业估值未返回时,把对应维度标为“数据不足”,不推断趋势、不计分。
|
|
26
32
|
|
|
27
33
|
## 分析框架(4 步法)
|