WOLF HOWL v2：自动化交易复盘与策略优化工具 - 数据驱动的每日分析

wolf-howl by senpi-ai/senpi-skills

124 周安装量

61 GitHub Stars

GitHub

安装命令

npx skills add https://github.com/senpi-ai/senpi-skills --skill wolf-howl

自动化数据分析金融科技

🇨🇳中文介绍

HOWL v2 — 狩猎、优化、赢利、学习

WOLF 终日狩猎。夜晚，它会 HOWL —— 回顾每一次捕杀与失手，磨砺其本能，以便明日以更敏锐的姿态醒来。

为 WOLF 策略提供数据驱动的自我改进建议的自动化每日复盘。

设置

运行设置脚本以配置夜间 HOWL：

python3 scripts/howl-setup.py --wallet {WALLET} --chat-id {CHAT_ID}

代理已知道钱包和聊天 ID —— 它只需要创建定时任务。可选设置运行时间（默认：本地时间 23:55）和时区。

工作原理

定时任务每日触发并生成一个独立的子代理，该子代理执行以下操作：

1. 收集数据

读取 memory/YYYY-MM-DD.md（今天 + 昨天）
读取 MEMORY.md 获取累积上下文
读取所有 dsl-state-WOLF-*.json 文件（active = 当前持仓，inactive = 已平仓交易）
读取 wolf-strategy.json 获取当前配置
读取 wolf-trade-counter.json 获取 FDR 数据（v5.1+）

广告位招租

在这里展示您的产品或服务

触达数万 AI 开发者，精准高效

联系我们

2. 分析每笔已平仓交易

针对每笔交易：资产、方向、入场/出场价格、盈亏（毛额和净额）、ROE、支付的手续费、持仓时长、最大 ROE（最高水位）、达到的 DSL 层级、入场信号类型和质量（理由数量、排名跃升、贡献速度、交易者数量）、入场时与出场时的 SM 信念度、平仓触发原因（DSL 突破/阶段 1 自动止损/停滞/信念度崩溃/轮换/手动）。

胜率、平均盈利 vs 平均亏损、盈利因子（毛额 AND 净额）
支付的总手续费、手续费拖累比率（手续费 / 账户价值）
净盈亏 vs 毛盈亏 —— 如果毛额为正但净额为负，手续费是问题所在

信号质量相关性（更高的理由数量 → 更好的结果？）
FIRST_JUMP vs IMMEDIATE_MOVER vs DEEP_CLIMBER 的胜率
入场排名 vs 结果 —— 排名 #20-30 的入场是否优于 #10-20？

层级分布（有多少达到第 1/2/3/4 层 vs 阶段 1 平仓？）
阶段 1 自动止损次数和节省金额（如果没有止损，它们会损失多少？）

持仓时段分桶分析（v2 — 关键）：

< 30 分钟：次数、胜率、盈亏、手续费（预期：差 —— 轮换和恐慌性止损）
30-90 分钟：次数、胜率、盈亏、手续费（预期：最佳区间）
90+ 分钟：次数、胜率、盈亏、手续费（预期：无效持仓）

方向分析（v2）：

LONG 胜率、盈亏、盈利因子
SHORT 胜率、盈亏、盈利因子
标记任一方向表现显著更差的情况（市场状态不匹配）

仓位利用率（占用时间 vs 空置时间的百分比）
无效持仓时长（亏损单在止损前坐了多久）
错失的机会（我们未交易的最大涨幅标的）
轮换次数和净成本（轮换交易支付的手续费）

区分盈利交易与亏损交易的入场模式
DSL 有效性（太紧？太松？）
停滞阈值（是否足够快地止损无效持仓？）
扫描器筛选器准确性（捕捉到正确的信号了吗？）
仓位规模优化
时机模式
手续费拖累模式（v2）：代理是否过度交易？在 FDR 达到 10% 之前进行了多少次交易？
持仓时段模式（v2）：短期持仓交易是否系统性更差？
方向偏差（v2）：代理是否在逆势交易？在抛售中做多 = 灾难。
巨额交易依赖（v2）：少数几笔大额盈利交易是否拯救了局面？多少交易产生了 80% 的盈亏？

将完整报告保存到 memory/howl-YYYY-MM-DD.md，包含：

摘要统计（交易次数、胜率、毛盈亏、净盈亏、手续费、FDR、毛/净盈利因子）
交易日志表格
持仓时段分桶明细
方向明细（LONG vs SHORT）
有效/无效之处（数据支持）
模式洞察
推荐改进措施（高/中/低置信度）
配置更改建议

6. 更新记忆并交付

将提炼的摘要追加到 MEMORY.md
向用户发送简洁的 Telegram 摘要
漂移检查（v2）：如果连续 3+ 次 HOWL 建议相同的更改，则在报告中提升紧迫性

请参阅 references/report-template.md 了解确切的输出格式。

v2 分析新增内容

这些是通过在实时 WOLF v5 交易数据上运行 HOWL v1 并发现盲点而得到的：

手续费拖累比率分析

来自第一次 HOWL 的最大洞察：手续费吞噬了整整一个盈利日。 32 笔交易 × 平均 $32 = $1,034 手续费（占账户的 18.3%）。毛盈亏为 +$888，但净盈亏为 -$146。HOWL 必须计算并在盈亏旁边显著显示 FDR。如果毛盈利因子 > 1.0 但净盈利因子 < 1.0，建议是进行更少、更高质量的交易 —— 而不是更好的入场。

持仓时段分桶分析

按持仓时间分桶的交易显示，< 30 分钟的交易系统性糟糕（合计 -$705），而 60-90 分钟的交易是最佳区间（+$704，57% 胜率）。HOWL 必须对每笔交易进行分桶，并标记 < 30 分钟的交易是否为负贡献者。

LONG 交易 4 胜 12 负（25% 胜率，盈利因子 0.05） vs 盈利的 SHORT 交易。HOWL 必须按方向拆分指标，并在某一方向表现显著不佳时标记 —— 这表明存在状态不匹配（在抛售中交易 LONG）。

巨额交易依赖性

3 笔第 4 层交易产生了 +$1,443，而其他所有交易合计为 -$556。HOWL 必须识别盈亏的百分之多少来自前 N 笔交易，以及策略在没有这些交易的情况下是否还能存活。

每次轮换成本约为 $65（平仓费 + 开仓费）。HOWL 必须追踪轮换次数、总轮换成本，以及轮换是否产生了净正结果。

每项建议必须有过去 24 小时数据的支持
不要改变有效的东西 —— 如果胜率高，就不要修复没有坏掉的东西
小的渐进式改进 > 大的全面改革
如果 24 小时内交易 < 3 笔，仍需分析“为什么这么少？” —— 不要直接跳过。是 FDR 门锁定了吗？市场死气沉沉？信号筛选过于激进？
比较今日数据与累积统计数据以发现趋势
残酷地诚实 —— 不要粉饰损失
始终区分毛额与净额。 一个在扣除手续费后亏损的“盈利”策略并非真正盈利。
及早标记状态不匹配。 如果 LONG 交易 0 胜 5 负，就说“停止做多 LONG” —— 不要等到 0 胜 12 负。

编辑 references/analysis-prompt.md 以调整子代理分析的内容。该提示词在运行时被子代理读取，因此更改会在下一次 HOWL 时生效，无需重启定时任务。

文件	用途
`scripts/howl-setup.py`	设置向导 —— 创建夜间 HOWL 定时任务
`references/analysis-prompt.md`	完整的子代理分析提示词（可编辑）
`references/report-template.md`	输出报告格式

MIT — 由 Senpi (https://senpi.ai) 构建。源代码：https://github.com/Senpi-ai/senpi-skills

🇺🇸English

HOWL v2 — Hunt, Optimize, Win, Learn

The WOLF hunts all day. At night, it HOWLs — reviewing every kill and miss, sharpening its instincts, and waking up sharper tomorrow.

Automated daily retrospective with data-driven self-improvement suggestions for the WOLF strategy.

Setup

Run the setup script to configure the nightly HOWL:

python3 scripts/howl-setup.py --wallet {WALLET} --chat-id {CHAT_ID}

The agent already knows wallet and chat ID — it just needs to create the cron. Optionally set run time (default: 23:55 local) and timezone.

How It Works

The cron fires daily and spawns an isolated sub-agent that:

1. Gathers Data

Reads memory/YYYY-MM-DD.md (today + yesterday)
Reads MEMORY.md for cumulative context
Reads all dsl-state-WOLF-*.json files (active = current positions, inactive = closed trades)
Reads wolf-strategy.json for current config
Reads wolf-trade-counter.json for FDR data (v5.1+)
Queries Senpi trade history via mcporter
Reads scanner script for current filter thresholds

2. Analyzes Each Closed Trade

For every trade: asset, direction, entry/exit price, PnL (gross and net), ROE, fees paid, duration, max ROE (high water), DSL tier reached, entry signal type and quality (reason count, rank jump, contrib velocity, trader count), SM conviction at entry vs exit, close trigger (DSL breach/Phase 1 auto-cut/stagnation/conviction collapse/rotation/manual).

3. Computes Metrics

Core:

Win rate, avg winner vs avg loser PnL, profit factor (gross AND net)
Total fees paid, fee drag ratio (fees / account value)
Net PnL vs gross PnL — if gross is positive but net is negative, fees are the problem

Signal Quality:

Signal quality correlation (do higher reason counts → better outcomes?)
FIRST_JUMP vs IMMEDIATE_MOVER vs DEEP_CLIMBER win rates
Entry rank vs outcome — are entries from rank #20-30 better than #10-20?

DSL Performance:

Tier distribution (how many reached Tier 1/2/3/4 vs Phase 1 closes?)
Phase 1 auto-cut count and savings (what would they have lost without the cut?)

Holding Period Buckets (v2 — critical):

< 30 min: count, win rate, PnL, fees (expect: poor — rotations and panic cuts)
30-90 min: count, win rate, PnL, fees (expect: sweet spot)
90+ min: count, win rate, PnL, fees (expect: dead weight)

Direction Analysis (v2):

LONG win rate, PnL, profit factor
SHORT win rate, PnL, profit factor
Flag if one direction is dramatically worse (regime mismatch)

Other:

Slot utilization (% time filled vs empty)
Dead weight duration (how long losers sat before cut)
Missed opportunities (top movers we didn't trade)
Rotation count and net cost (fees paid for rotation trades)

4. Identifies Patterns

Entry patterns distinguishing winners from losers
DSL effectiveness (too tight? too loose?)
Stagnation thresholds (cutting dead weight fast enough?)
Scanner filter accuracy (catching right signals?)
Position sizing optimization
Timing patterns
Fee drag pattern (v2): Is the agent overtrading? How many trades before FDR hits 10%?
Holding period pattern (v2): Are short-hold trades systematically worse?
Direction bias (v2): Is the agent fighting the trend? LONGs in a selloff = disaster.
Monster trades (v2): Did a few big winners save the day? How many trades produced 80% of PnL?

5. Produces Report

Saves full report to memory/howl-YYYY-MM-DD.md with:

Summary stats (trades, win rate, gross PnL, net PnL, fees, FDR, profit factor gross/net)
Trade log table
Holding period buckets breakdown
Direction breakdown (LONG vs SHORT)
What worked / what didn't (data-backed)
Pattern insights
Recommended improvements (high/medium/low confidence)
Config change suggestions

6. Updates Memory & Delivers

Appends distilled summary to MEMORY.md
Sends concise Telegram summary to user
Drift check (v2): If 3+ consecutive HOWLs suggest the same change, escalate urgency in the report

Report Format

See references/report-template.md for the exact output format.

v2 Analysis Additions

These were discovered by running HOWL v1 on live WOLF v5 trading data and finding blind spots:

Fee Drag Ratio Analysis

The single biggest insight from the first HOWL: fees ate an entire profitable day. 32 trades × $32 avg = $1,034 in fees (18.3% of account). Gross PnL was +$888, but net was -$146. HOWL must compute and prominently display FDR alongside PnL. If gross PF > 1.0 but net PF < 1.0, the recommendation is fewer, higher-quality trades — not better entries.

Holding Period Buckets

Trades bucketed by hold time revealed that < 30 min trades were systematically terrible (-$705 combined) while 60-90 min trades were the sweet spot (+$704, 57% WR). HOWL must bucket every trade and flag if < 30 min trades are negative contributors.

Direction Regime Detection

4W/12L on LONGs (25% WR, PF 0.05) vs profitable SHORTs. HOWL must split metrics by direction and flag when one side is dramatically underperforming — this indicates a regime mismatch (trading LONGs in a selloff).

Monster Trade Dependency

3 Tier 4 trades produced +$1,443 while everything else combined was -$556. HOWL must identify what % of PnL came from top N trades and whether the strategy would survive without them.

Rotation Cost Tracking

Each rotation costs ~$65 (close fee + open fee). HOWL must track rotation count, total rotation cost, and whether rotations produced net positive outcomes.

Rules

Every recommendation must be backed by data from the last 24h
Don't change things that are working — if win rate is high, don't fix what isn't broken
Small incremental improvements > big overhauls
If < 3 trades in 24h, still analyze "why so few?" — don't just skip. Was the FDR gate locked? Was the market dead? Were signals filtered too aggressively?
Compare today vs cumulative stats to spot trends
Be brutally honest — no sugarcoating losses
Always separate gross vs net. A "profitable" strategy that loses money after fees is not profitable.
Flag regime mismatches early. If LONGs are 0-for-5, say "stop taking LONGs" — don't wait for 0-for-12.

Customization

Edit references/analysis-prompt.md to adjust what the sub-agent analyzes. The prompt is read by the sub-agent at runtime, so changes take effect on the next HOWL without restarting crons.

Files

File	Purpose
`scripts/howl-setup.py`	Setup wizard — creates the nightly HOWL cron
`references/analysis-prompt.md`	Full sub-agent analysis prompt (editable)
`references/report-template.md`	Output report format

License

MIT — Built by Senpi (https://senpi.ai). Source: https://github.com/Senpi-ai/senpi-skills

Weekly Installs

111

Repository

senpi-ai/senpi-skills

GitHub Stars

First Seen

Feb 27, 2026

Security Audits

Gen Agent Trust HubPass SocketWarn SnykFail

Installed on

openclaw108

opencode103

gemini-cli103

github-copilot103

codex103

kimi-cli103

Python PDF处理教程：合并拆分、提取文本表格、创建PDF文件

65,000 周安装