Calibration des probabilités LLM IA + Dimensionnement Half-Kelly : intégration étape par étape de l’API DeepSeek / OpenAI / Ollama pour des ordres +EV

# 0. 安装依赖并配置大模型 API 密钥 (支持 DeepSeek / OpenAI / Claude / 本地 Ollama)
# pip install openai requests
# export LLM_API_KEY="sk-你的大模型API密钥"
# export LLM_BASE_URL="https://api.deepseek.com/v1"  # 或 https://api.openai.com/v1 或本地 http://127.0.0.1:11434/v1

import os, json, math
from openai import OpenAI

# 1. 初始化大模型客户端 (兼容所有 OpenAI 格式接口)
llm_client = OpenAI(
    api_key=os.environ.get("LLM_API_KEY", "sk-your-api-key"),
    base_url=os.environ.get("LLM_BASE_URL", "https://api.deepseek.com/v1"),
)

def get_llm_catalyst_likelihood(market_question: str, resolution_rules: str, breaking_news: str) -> dict:
    """调用大模型 API 分析突发新闻对盘口结算的边际影响，强制返回 JSON 格式似然比 (0.50 ~ 2.00)"""
    system_prompt = (
        "You are a quantitative prediction-market analyst. Compare the breaking news against the exact "
        "UMA resolution rules. Output ONLY valid JSON: "
        '{"likelihood_ratio": float_between_0.5_and_2.0, "confidence": float_0_to_1, "reason": "short summary"} '
        "where 1.0 = neutral noise, >1.15 = bullish YES catalyst, <0.85 = bearish NO catalyst."
    )
    user_prompt = f"Market: {market_question}\nRules: {resolution_rules}\nBreaking News: {breaking_news}"
    resp = llm_client.chat.completions.create(
        model=os.environ.get("LLM_MODEL", "deepseek-chat"),  # 或 gpt-4o-mini / qwen-plus / gemma4:31b
        messages=[
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": user_prompt},
        ],
        response_format={"type": "json_object"},
        temperature=0.1,
    )
    return json.loads(resp.choices[0].message.content)

def bayesian_update_prob(prior_mkt_prob: float, whale_net_flow_usd: float, llm_likelihood_ratio: float) -> float:
    """融合盘口先验概率、Polygon 巨鲸净流入与 AI 催化剂似然比，输出校准后真实胜率 p*"""
    p0 = max(0.01, min(0.99, prior_mkt_prob))
    log_odds = math.log(p0 / (1.0 - p0))
    whale_shift = max(-0.45, min(0.45, whale_net_flow_usd / 150000.0))
    llm_shift = math.log(max(0.5, min(2.0, llm_likelihood_ratio)))
    post_odds = math.exp(log_odds + whale_shift + llm_shift)
    return round(post_odds / (1.0 + post_odds), 4)

# 2. 实战串联：调用大模型 -> 贝叶斯校准 p* -> 计算半凯利最优仓位
p_mkt = 0.35
ai_res = get_llm_catalyst_likelihood(
    market_question="Will the Fed cut rates by 25 bps at the next FOMC meeting?",
    resolution_rules="Resolves YES if official FOMC statement lowers target federal funds rate.",
    breaking_news="US Core CPI prints 0.1% below consensus; two Fed governors signal dovish shift.",
)
lr = float(ai_res.get("likelihood_ratio", 1.0))
p_star = bayesian_update_prob(prior_mkt_prob=p_mkt, whale_net_flow_usd=85000, llm_likelihood_ratio=lr)
half_kelly = max(0.0, 0.5 * ((p_star - p_mkt) / (1.0 - p_mkt)))
print(f"[AI 诊断] 似然比={lr:.2f} ({ai_res.get('reason')}) -> 校准胜率 p*={p_star:.1%} | 建议半凯利仓位={half_kelly:.1%}")