Principles (MUST follow, read this FIRST)

• NEVER ask the user for parameters. User says "茅台均线能赚钱吗" → you pick strategy=ema-crossover, symbol=600519.SH, interval=D, and run it. Use sensible defaults for everything.

• Resolve stock names yourself. "茅台" = 600519.SH, "平安" = 601318.SH. Use catquant.resolve.resolve(query) to map names to codes. NEVER ask the user "请问茅台的代码是什么".

• Validate data BEFORE running backtest — NEVER skip this. Call check_available(code) and if it returns False , STOP immediately and tell the user. Do NOT assume data exists. Do NOT create a script without checking first. Most stocks are NOT covered by the free API.

• Handle errors gracefully. If data source fails, suggest alternatives (TDX local data, or contact developer). Never show raw Python tracebacks to the user.

• One sentence = one complete result. The user's single sentence should produce a full backtest report with charts. No intermediate questions, no "请确认参数" dialogues.

Setup

The catquant package lives inside this skill directory ($SKILL_DIR ). Every generated script must:

import sys, os sys.path.insert(0, os.path.join(os.path.dirname(file), "$SKILL_DIR")) from dotenv import load_dotenv, find_dotenv load_dotenv(find_dotenv())

First run: pip install -r $SKILL_DIR/requirements.txt if dependencies are missing.

Step 0: Resolve stock and validate data (MANDATORY — do NOT skip)

STOP. You MUST complete BOTH steps below BEFORE writing any backtest or scan script. If you skip check_available , the script WILL fail and the user gets nothing.

from catquant.resolve import resolve, check_available

Step 0a: Resolve name to code

code, name = resolve("茅台") # -> ("600519.SH", "茅台")

Step 0b: Check data — YOU MUST DO THIS, DO NOT ASSUME DATA EXISTS

ok, source, hint = check_available(code) if not ok: print(hint) # Tell the user WHY it failed and HOW to fix it sys.exit(1) # STOP HERE. Do NOT create a backtest script.

CRITICAL: If check_available returns (False, ...) you MUST:

• STOP — do NOT proceed to create or run any backtest script

• Tell the user the stock is not covered, in plain language

• Give actionable next steps:

• If TDX_DIR is set but stock not found → "Please download this stock's data in your TDX client"

• If TDX_DIR is not set → "Set up TDX local data, or contact https://www.jjmfc.com for full-market access"

• Do NOT silently fail, do NOT produce an empty backtest, do NOT skip this check

The free FaceCat API only covers ~84 stocks. Most stocks (including 茅台、比亚迪、宁德时代 etc.) are NOT covered. You MUST verify before proceeding.

If ok is True: use the returned source value in get_history() :

• source="facecat" (default) or source="tdx", tdx_dir=os.environ["TDX_DIR"]

Command: backtest

Parse $ARGUMENTS : backtest [strategy] [symbol] [interval]

• strategy: ema-crossover, rsi, macd, kdj, boll. Default: ema-crossover (do NOT ask)

• symbol: stock name or code. Default: SH600000 (do NOT ask)

• interval: D, 5m, 1m. Default: D (do NOT ask)

Steps:

• Resolve stock name → code (Step 0 above)

• Validate data availability (Step 0 above)

• Read catquant-expert for API and strategy patterns

• Create script at backtesting/{strategy_name}/{symbol}_{strategy}_backtest.py

• Script must: load data, compute signals, run backtest, export JSON, render charts with indicator lines, print metrics

• Run the script

• Explain the backtest report in plain language to the user

Chart Rendering

Always render charts with the strategy's indicator lines. Use overlays for lines on the price chart and panels for indicator subplots:

Strategy overlays (on price) panels (subplots)

EMA Cross EMA fast + EMA slow lines MACD panel (DIF/DEA + MACD bars + zero_line)

MACD EMA(12) + EMA(26) lines MACD panel (DIF/DEA + MACD bars + zero_line)

RSI EMA(14) line RSI panel (RSI line)

KDJ

KDJ panel (K/D/J lines)

BOLL Upper + Mid + Lower lines + fill

Custom Indicator lines used in signals Relevant oscillator

Example:

render(result, bars, outdir, "kline", overlays=[ {"data": ema5, "label": "EMA5", "color": "#ff9800"}, {"data": ema20, "label": "EMA20", "color": "#2196f3"}, ], panels=[{ "title": "MACD", "lines": [ {"data": dif, "label": "DIF", "color": "#2962ff"}, {"data": dea, "label": "DEA", "color": "#ff6d00"}, ], "bars": [{"data": macd, "label": "MACD"}], "zero_line": True, }]) render(result, bars, outdir, "equity")

Command: scan

Parse $ARGUMENTS : scan [screening criteria]

Natural language criteria. Default: scan all stocks (do NOT ask for clarification if criteria is clear enough).

Steps:

• Parse criteria into pre_filter (PriceData fields) and filter_fn (K-line indicators)

• Create script at scanning/{name}_scan.py

• Run and show results

Scan types

• Price/volume/PE only: quick_scan(pre_filter) — instant, no K-line download

• Needs indicators: scan(filter_fn, pre_filter) — two-layer, downloads K-lines

pre_filter (Layer 1)

lambda p: p.volume > 0 and p.close > 3 and "ST" not in p.name

Fields: p.code, p.name, p.close, p.open, p.high, p.low, p.volume, p.amount, p.lastClose, p.pe, p.totalShares, p.flowShares, p.upperLimit, p.lowerLimit

filter_fn (Layer 2)

def filter_fn(code, name, bars): if len(bars) < 30: return None close, high, low, vol = bars_to_arrays(bars) # compute indicators... if condition: return {"score": value, "reason": "..."} return None

API

scan(filter_fn, pre_filter=None, universe=None, count=250, cycle=1440, source="facecat", max_results=20, sort_key="score", ascending=False, verbose=True, refresh=False)

quick_scan(pre_filter=None, max_results=50, sort_key="close", ascending=False)

PDF Reports — Full Spec (when user asks for a report)

When the user requests a PDF report, generate a complete, high-quality report — not a minimal one. The report must include AI commentary that interprets results in plain language, not just numbers.

Page Structure

Page 1 — Cover Page 2 — Executive Summary (metrics + AI verdict) Page 3 — K-line Chart (embed kline.png full-width) Page 4 — Equity Curve (embed equity.png full-width) Page 5 — Trade Records (table, paginated if > 30 trades) Page 6 — Risk Analysis (risk ratios + AI risk commentary) Page 7 — Conclusion & Suggestions

Chinese Font Registration (MANDATORY — do this first)

import sys from reportlab.pdfbase import pdfmetrics from reportlab.pdfbase.ttfonts import TTFont

def _register_cjk_font(): candidates = [] if sys.platform == "win32": candidates = [ r"C:\Windows\Fonts\simsun.ttc", r"C:\Windows\Fonts\msyh.ttc", r"C:\Windows\Fonts\simhei.ttf", ] elif sys.platform == "darwin": candidates = [ "/System/Library/Fonts/PingFang.ttc", "/Library/Fonts/Arial Unicode MS.ttf", ] else: candidates = [ "/usr/share/fonts/truetype/wqy/wqy-microhei.ttc", "/usr/share/fonts/opentype/noto/NotoSansCJK-Regular.ttc", ] for path in candidates: try: pdfmetrics.registerFont(TTFont("CJK", path)) return "CJK" except Exception: continue raise RuntimeError("No CJK font found. Windows: simsun.ttc; Linux: sudo apt install fonts-wqy-microhei")

CJK_FONT = _register_cjk_font()

Never use default Helvetica or Times-Roman for Chinese — they produce garbled output.

Style Setup

from reportlab.lib.styles import getSampleStyleSheet, ParagraphStyle from reportlab.lib import colors

styles = getSampleStyleSheet()

def cn(size=10, bold=False, color=colors.black): return ParagraphStyle("cn", fontName=CJK_FONT, fontSize=size, leading=size * 1.5, textColor=color, fontWeight="Bold" if bold else "Normal")

title_style = cn(22, bold=True) h2_style = cn(13, bold=True, color=colors.HexColor("#1a237e")) body_style = cn(10) caption_style = cn(9, color=colors.HexColor("#555555")) verdict_style = cn(11, bold=True)

Page 1 — Cover

[股票名称 + 代码] (28pt, bold, centered) [策略名称] (16pt, centered) [回测区间: YYYY-MM-DD ~ YYYY-MM-DD] [生成时间: YYYY-MM-DD HH:MM] [MaoQuant · 量化回测报告] (footer)

Page 2 — Executive Summary

Metrics table — use a 2-column grid layout:

指标 值

初始资金 100,000 元

最终资产 X,XXX,XXX 元

总收益率 +XX.XX%

年化收益率 +XX.XX%

最大回撤 -XX.XX%

夏普比率 X.XX

索提诺比率 X.XX

卡玛比率 X.XX

总交易次数 XX 次

胜率 XX.XX%

盈亏比 X.XX

平均持仓 XX 天

最长连续亏损 X 次

AI 综合评语 — generate from metrics using the rules below.

AI Commentary Rules

Generate natural-language commentary based on these thresholds. Write in Chinese, conversational tone, 3-5 sentences per section.

Overall Verdict (one of four tones):

def overall_verdict(m): tr = m["total_return"] md = abs(m["max_drawdown"]) sr = m["sharpe_ratio"] if tr > 0.3 and md < 0.2 and sr > 1.0: return "strong" # "策略表现出色，..." elif tr > 0 and sr > 0.5: return "decent" # "策略整体有效，..." elif tr > 0: return "marginal" # "策略勉强盈利，但风险控制有待加强..." else: return "losing" # "策略在此区间亏损，..."

Commentary templates (customize with actual numbers):

strong: "策略在{years}年回测中表现出色，累计收益{tr:.1%}，年化{ar:.1%}。 最大回撤控制在{md:.1%}以内，夏普比率{sr:.2f}显示风险调整后收益优秀。 适合作为核心策略参考。"

decent: "策略整体有效，累计盈利{tr:.1%}，但存在一定波动。 最大回撤{md:.1%}在可接受范围内，夏普比率{sr:.2f}处于合理水平。 建议配合止损策略使用。"

marginal: "策略勉强盈利{tr:.1%}，但胜率仅{wr:.1%}，盈亏比{pf:.2f}偏低。 最大回撤{md:.1%}较大，持仓期间心理压力较高。 建议优化参数或换用其他策略。"

losing: "策略在此区间亏损{tr:.1%}，共{total}笔交易，胜率{wr:.1%}。 最大回撤达{md:.1%}，说明策略与该股票当前趋势不匹配。 建议尝试其他策略或等待趋势确认。"

Win rate interpretation:

• win_rate > 0.6 and profit_factor < 1.2 → "胜率较高但盈亏比偏低，可能存在小赚大亏的问题，注意设置止盈。"

• win_rate < 0.4 and profit_factor > 2.0 → "胜率较低但盈亏比优秀，属于趋势跟踪型策略，需要心理承受能力。"

• max_consecutive_losses > 5 → "最长连续亏损{n}次，实盘时需做好心态管理。"

Risk rating (show as colored badge):

def risk_rating(m): md = abs(m["max_drawdown"]) if md < 0.15: return ("低风险", "#4caf50") elif md < 0.25: return ("中等风险", "#ff9800") elif md < 0.40: return ("较高风险", "#f44336") else: return ("高风险", "#b71c1c")

Page 5 — Trade Records Table

Columns: 序号 | 买入日期 | 买入价 | 卖出日期 | 卖出价 | 股数 | 手续费 | 盈亏(元) | 持仓天数

• Profitable rows: light green background (#e8f5e9 )

• Loss rows: light red background (#ffebee )

• Open positions: light yellow (#fffde7 ) + "持仓中" in 卖出日期 column

• If > 30 trades, show first 15 + last 15 with a "…共 N 笔交易" separator

from reportlab.platypus import TableStyle from reportlab.lib import colors

def trade_row_style(trades): styles = [] for i, t in enumerate(trades): row = i + 1 # header is row 0 if t["exit_date"] == "": bg = colors.HexColor("#fffde7") elif t["pnl"] > 0: bg = colors.HexColor("#e8f5e9") else: bg = colors.HexColor("#ffebee") styles.append(("BACKGROUND", (0, row), (-1, row), bg)) return styles

Page 6 — Risk Analysis

Display these ratios with plain-language explanation next to each:

指标 值 含义

夏普比率 X.XX

1.0 表示风险调整后收益良好

索提诺比率 X.XX 仅考虑下行波动，> 1.0 为佳

卡玛比率 X.XX 年化收益/最大回撤，> 0.5 可接受

最大回撤 -XX% 最大亏损幅度

最长回撤持续 XX 天 深度亏损持续时间

AI 风险评语: 2-3 sentences interpreting the risk profile:

"该策略最大回撤{md:.1%}，最长持续{dur}个交易日。 卡玛比率{calmar:.2f}，说明每承受1%的最大回撤， 可获得{calmar:.2f}%的年化收益。{rating}。"

Page 7 — Conclusion & Suggestions

Always include these 4 sections:

• 策略总结 — one paragraph summarizing what the strategy does and how it performed

• 优势 — bullet list of strong points (high win rate / controlled drawdown / good Sharpe / etc.)

• 风险提示 — bullet list of weaknesses (high drawdown / low win rate / short sample / etc.)

• 改进建议 — concrete suggestions:

• If avg_holding_bars < 5 : "持仓过短，频繁交易产生较高手续费，建议提高信号阈值"

• If max_drawdown < -0.30 : "最大回撤偏大，建议加入止损 stop_loss=0.08 "

• If total_trades < 10 : "样本量不足，建议延长回测区间或换用更频繁触发的策略"

• If win_rate < 0.35 : "胜率较低，建议配合趋势过滤条件（如均线方向）"

Output Path

Save to: backtesting/{strategy}/{symbol}_report.pdf

TDX Formula Translation (通达信公式转 Python)

When the user pastes a TDX/通达信 formula and asks to backtest or scan it, translate it to Python using these rules. Do NOT run the formula text directly — always convert to Python first.

TDX → Python mapping

TDX Python (bars = List[SecurityData], arrays from bars_to_arrays )

C / O / H / L / V

close[i] / open_[i] / high[i] / low[i] / vol[i]

REF(X, N)

x[i - N]

HHV(X, N)

np.max(x[i-N+1:i+1])

LLV(X, N)

np.min(x[i-N+1:i+1])

MA(X, N)

np.mean(x[i-N+1:i+1])

EMA(X, N)

ema_series(x, N)[i]

CROSS(A, B)

cross_above(a, b)[i]

AND / OR / NOT

and / or / not (Python booleans)

:= (intermediate variable) var = expr

XG:A AND B AND C; (final signal) signal = a and b and c

Sliding-window scan pattern

For a formula that checks the last N bars at each point in history:

def check_pattern(o, h, l, c, v, i): """Check if pattern triggers at bar index i.""" if i < 5: # guard: need enough lookback return False # translate each TDX condition directly: # A1:=REF(C,3)>REF(O,3) -> a1 = c[i-3] > o[i-3] # A2:=REF(V,3)>REF(V,4)*1.8 -> a2 = v[i-3] > v[i-4] * 1.8 # ... return a1 and a2 # and all other conditions

Scan all bars

signals = [i for i in range(5, len(bars)) if check_pattern(o, h, l, c, v, i)]

Formula backtest vs single-stock backtest

• User says "回测这个公式" (no specific stock) → Do a historical pattern scan across full TDX universe. For each signal found, record entry price and return after N days. Output: win rate, avg return, profit factor, trade list CSV.

• User says "用这个公式回测某只股票" → Single-stock backtest.run() with buy_signal / sell_signal arrays derived from the formula.

For historical pattern scans (全市场历史扫描)

Build universe from TDX vipdoc

def get_tdx_universe(tdx_dir): universe = [] for mkt, pfx in [("sh","SH"), ("sz","SZ")]: lday = os.path.join(tdx_dir, "vipdoc", mkt, "lday") for f in os.listdir(lday): if not f.endswith(".day"): continue num = f[2:-4] # A-shares only: skip indices/funds/bonds if mkt == "sh" and not (num.startswith("6") or num.startswith("688")): continue if mkt == "sz" and not (num.startswith("0") or num.startswith("3")): continue universe.append(f"{num}.{pfx}") return universe

Per-stock: slide window over full history

for code in universe: bars = get_history(code, count=2000, cycle=1440, source="tdx", tdx_dir=TDX_DIR) o, h, l, c, v = ... # arrays for i in range(lookback, len(bars)): if check_pattern(o, h, l, c, v, i): entry = c[i] for hold in [5, 10, 20]: if i + hold < len(bars): ret = (c[i+hold] - entry) / entry * 100 record(code, date, entry, ret, hold)

涨停 detection

def is_limit_up(code, close, prev_close): from catquant.backtest import get_limit_pct limit = get_limit_pct(code) return close >= round(prev_close * (1 + limit) - 0.01, 2)

Output stats after pattern scan

Always report:

• 触发次数、按年份分布

• 持有 5/10/20 天的：胜率、平均收益、盈亏比、最大单次盈亏

• Top 10 最佳触发案例（表格）

• 优化建议（加大盘过滤/收紧阈值/止损）

Matplotlib Chinese Font (图表中文显示)

Matplotlib 默认字体不支持中文，所有含中文标签/标题的图表必须先注册字体，否则显示方块乱码。

import matplotlib import matplotlib.pyplot as plt from matplotlib.font_manager import fontManager

def _setup_mpl_cjk(): candidates = [] import sys, os if sys.platform == "win32": candidates = [ r"C:\Windows\Fonts\simsun.ttc", r"C:\Windows\Fonts\msyh.ttc", r"C:\Windows\Fonts\simhei.ttf", ] elif sys.platform == "darwin": candidates = ["/System/Library/Fonts/PingFang.ttc"] else: candidates = ["/usr/share/fonts/truetype/wqy/wqy-microhei.ttc"] for path in candidates: if os.path.exists(path): fontManager.addfont(path) from matplotlib.font_manager import FontProperties name = FontProperties(fname=path).get_name() matplotlib.rcParams["font.sans-serif"] = [name, "DejaVu Sans"] matplotlib.rcParams["axes.unicode_minus"] = False return # fallback: suppress warnings and use English labels import warnings warnings.filterwarnings("ignore", "Glyph.*missing from font")

_setup_mpl_cjk() # call once before any plt usage

Rules:

• Call _setup_mpl_cjk() at module level, before any plt. call

• If no CJK font found, fall back to English axis labels (never crash)

• axes.unicode_minus = False prevents minus sign rendering as a box

Constraints

• Raw data never enters AI context — use catquant.data_engine inside scripts only

• A-share compliance auto-enforced by backtest.run() :

• T+1, price limits (10%/20%/30%), lot sizing (100 shares), fees

• Scripts go in backtesting/ or scanning/

• No emoji in code or output

• Charts via catquant.chart.render()

• PDF reports must use a registered CJK font (see PDF Reports section above)

Reference

• catquant-expert — API, indicators, strategy patterns

• data — Data engine, SecurityData, symbol format