Polymarket NLP 情感套利系统

# nlp/sentiment_engine.py — System Prompt 设计（Chain-of-Thought 压缩）

SYSTEM_PROMPT = """
You are a prediction market analyst. Analyze breaking news for Polymarket relevance.
Respond ONLY with valid JSON. No markdown, no explanation outside JSON.

JSON schema (required):
{
  "relevance": <float 0-1>,
  "direction": <"YES"|"NO"|"NEUTRAL">,
  "prob_delta": <float -1 to 1>,
  "confidence": <float 0-1>,
  "category": <"fed_rate"|"us_election"|"geopolitical"|"crypto_reg"|"macro"|"political"|"other">,
  "reasoning": <string max 60 chars>
}

Rules:
- relevance: how likely this news affects an active Polymarket prediction market
- direction: YES = increases probability of YES outcome; NO = decreases it
- prob_delta: estimated absolute change to event probability (e.g. 0.15 = +15%)
- confidence: your certainty in this assessment
- If news is irrelevant, set relevance<0.3, direction=NEUTRAL, prob_delta=0
"""

# 示例 LLM 输出（Fed降息新闻）：
# {
#   "relevance": 0.97,
#   "direction": "YES",
#   "prob_delta": 0.23,
#   "confidence": 0.91,
#   "category": "fed_rate",
#   "reasoning": "Emergency 50bps cut signals Fed pivot confirmed"
# }

# execution/order_executor.py — 核心 Alpha 捕获逻辑

async def process_signal(self, signal: TradingSignal) -> bool:
    # ── 1. 搜索匹配市场 ──────────────────────
    cat_kws = CATEGORY_KEYWORDS.get(signal.category, [])
    markets_raw = await self._clob.search_markets(cat_kws)
    markets = [_parse_market(m) for m in markets_raw]

    # ── 2. 计算每个市场的 edge ───────────────
    for market in markets[:5]:
        token_id = (market.yes_token_id if signal.direction == "YES"
                    else market.no_token_id)
        current_price = await self._clob.get_midpoint(token_id)

        # LLM 估算公允价 = 当前价 + 概率偏移
        llm_fair = min(max(current_price + signal.prob_delta, 0.01), 0.99)
        edge = llm_fair - current_price   # 正数 = 被低估

        if edge >= risk_cfg.min_edge_pct:    # ≥ 7¢ 才触发
            best_market = market
            best_edge = edge

    # ── 3. Kelly 定仓 ────────────────────────
    # f* = (b×p - q)/b × 0.25 (1/4 Kelly)
    b = (1 - price) / price
    kelly_f = (b * (price + best_edge) - (1 - price - best_edge)) / b * 0.25
    size_usdc = min(max_capital * kelly_f / 0.15, max_capital)

    # ── 4. 提交 FOK 市价单扫平低价筹码 ───────
    order_id = await self._clob.place_market_order(
        token_id=token_id,
        side="BUY",
        size_usdc=size_usdc,
        price=current_price * 1.03,  # 3% 滑点容忍
    )
    return order_id is not None

async def check_exit_conditions(self):
    for cid, pos in self._open_positions.items():
        current = await self._clob.get_midpoint(pos.token_id)
        # 收敛平仓：当前价格接近 LLM 公允价 ±3¢
        if abs(current - pos.fair_value) <= 0.03:
            await self._clob.place_market_order(..., side="SELL")
        # 过度反应平仓：超过公允价 5¢（散户 FOMO）
        elif current > pos.fair_value + 0.05:
            await self._clob.place_market_order(..., side="SELL")

# feeds/news_firehose.py — 关键词过滤器

class KeywordFilter:
    """白名单命中 AND 黑名单未命中 → 放行"""

    def __init__(self):
        self.whitelist = [w.lower() for w in feed_cfg.keyword_whitelist]
        self.blacklist = [b.lower() for b in feed_cfg.keyword_blacklist]
        self._seen: set[str] = set()   # 去重窗口（最大 50K 条）

    def passes(self, item: NewsItem) -> bool:
        if item.item_id in self._seen:
            return False       # 去重
        low = item.text.lower()
        if any(b in low for b in self.blacklist):
            return False       # 黑名单命中
        if not any(w in low for w in self.whitelist):
            return False       # 白名单未命中
        self._seen.add(item.item_id)
        return True

# 白名单 45+ 条关键词示例：
WHITELIST = [
    "rate cut", "rate hike", "fed funds", "interest rate",
    "election", "resign", "impeach", "indicted",
    "ceasefire", "invasion", "sanction", "nuclear",
    "approve", "reject", "veto", "passed", "signed",
    "bitcoin", "crypto", "sec", "etf approved",
    "gdp", "recession", "unemployment", "cpi",
    "breaking", "urgent", "flash", ...
]

# risk/exposure_manager.py — LLM 幻觉熔断器

def record_close(self, category: str, pnl_usdc: float):
    self._daily_pnl += pnl_usdc

    if pnl_usdc < 0:
        self._consecutive_losses += 1
    else:
        self._consecutive_losses = 0
        self._hallucination_scale = min(
            self._hallucination_scale * 1.1, 1.0
        )

    # 幻觉熔断器三级响应
    if self._consecutive_losses >= 5:
        self._halted_until = time.time() + 3600  # 暂停1小时
        # LLM 可能产生幻觉或模型过时，强制停止
    elif self._consecutive_losses >= 3:
        self._hallucination_scale = 0.5           # 资金降至50%档

def approve_trade(self, signal: TradingSignal):
    # 置信度分级资金上限
    pct = {1: 0.15, 2: 0.05}.get(signal.news_item.tier, 0.01)
    max_usdc = min(
        risk_cfg.total_capital_usdc * pct * self._hallucination_scale,
        150.0   # 单事件绝对上限 $150
    )
    return (not self.is_halted()) and max_usdc > 0, max_usdc