Auto-Coder使用GPT-4o完成：在用TabPFN这个模型构建一个预测未来3天涨跌的分类任务

news2025/6/12 18:38:02

通过akshare库，获取股票数据，并生成TabPFN这个模型可以识别、处理的格式，写一个完整的预处理示例，并构建一个预测未来 3 天股价涨跌的分类任务

用TabPFN这个模型构建一个预测未来 3 天股价涨跌的分类任务，进行预测并输出准确率

代码

import os
import sys
import pandas as pd
import akshare as ak
# try:
#     import akshare as ak
# except ImportError:
#     print("请先运行: pip install akshare")
#     sys.exit(1)

def fetch_stock_data(stock_code="000001", start_date="20150101", end_date=None):
    """
    获取股票历史行情数据，兼容不同akshare接口
    """
    print(f"DEBUG: akshare version: {ak.__version__}")
    print(f"DEBUG: fetch_stock_data params: symbol={stock_code}, start_date={start_date}, end_date={end_date}")
    try:
        # 优先尝试 stock_zh_a_daily
        df = ak.stock_zh_a_daily(symbol=stock_code, adjust="qfq")
        print("DEBUG: 使用 ak.stock_zh_a_daily 成功")
        print("DEBUG: Columns before rename:", df.columns.tolist())
        print("DEBUG: Head before rename:\n", df.head())
        # 若有 start_date/end_date，筛选
        if "date" in df.columns:
            df["date"] = pd.to_datetime(df["date"])
            if start_date:
                df = df[df["date"] >= pd.to_datetime(start_date)]
            if end_date:
                df = df[df["date"] <= pd.to_datetime(end_date)]
            df = df.sort_values("date").reset_index(drop=True)
    except Exception as e:
        print("ERROR: ak.stock_zh_a_daily() 调用异常，尝试 fallback 原接口")
        print(f"Exception: {e}")
        try:
            df = ak.stock_zh_a_hist(symbol=stock_code, period="daily", start_date=start_date, end_date=end_date, adjust="qfq")
            print("DEBUG: 使用 ak.stock_zh_a_hist 成功")
            print("DEBUG: Columns before rename:", df.columns.tolist())
            print("DEBUG: Head before rename:\n", df.head())
        except Exception as e2:
            print("ERROR: Exception occurred while fetching stock data by both interfaces!")
            print(f"Exception: {e2}")
            return pd.DataFrame()
    if df.empty:
        print("ERROR: Fetched DataFrame is empty! Check stock code, date range, or network/API issues.")
        return df
    # 兼容列名
    rename_map = {
        "日期": "date",
        "开盘": "open",
        "收盘": "close",
        "最高": "high",
        "最低": "low",
        "成交量": "volume",
        "成交额": "amount",
        "振幅": "amplitude",
        "涨跌幅": "pct_chg",
        "涨跌额": "chg",
        "换手率": "turnover"
    }
    for k in list(rename_map.keys()):
        if k not in df.columns:
            rename_map.pop(k)
    df = df.rename(columns=rename_map)
    if "date" in df.columns:
        df["date"] = pd.to_datetime(df["date"])
        df = df.sort_values("date").reset_index(drop=True)
    # 自动补充pct_chg列（涨跌幅），百分比格式
    if "pct_chg" not in df.columns and "close" in df.columns:
        df["pct_chg"] = df["close"].pct_change() * 100
    return df

def create_features_and_labels(df, n_past=5, n_future=3):
    """
    构造特征和标签，标签为未来3日涨跌（1=涨，0=跌/平）
    """
    feats = []
    labels = []
    for idx in range(n_past, len(df) - n_future):
        past_slice = df.iloc[idx-n_past:idx]
        # 特征: 过去n_past日的收盘价、涨跌幅、成交量
        feature = []
        feature += list(past_slice["close"].values)
        feature += list(past_slice["pct_chg"].values)
        feature += list(past_slice["volume"].values)
        # 未来n_future日的收盘价均值
        future_close_mean = df.iloc[idx:idx+n_future]["close"].mean()
        curr_close = df.iloc[idx-1]["close"]
        # 涨跌标签: 未来3日均值 > 当前收盘价 => 1，否则0
        label = 1 if future_close_mean > curr_close else 0
        feats.append(feature)
        labels.append(label)
    feats_df = pd.DataFrame(feats, columns=[
        f"close_t-{i}" for i in range(n_past,0,-1)
    ] + [
        f"pct_chg_t-{i}" for i in range(n_past,0,-1)
    ] + [
        f"volume_t-{i}" for i in range(n_past,0,-1)
    ])
    feats_df["label"] = labels
    return feats_df

def save_for_tabpfn(df, out_csv):
    """
    保存为TabPFN模型可读取的csv格式
    """
    df.to_csv(out_csv, index=False)
    print(f"已保存至: {out_csv}")
    print(df.head())

def main():
    # 拉取A股代码表，打印前10条，辅助判断symbol格式
    print("尝试拉取A股代码表，辅助symbol格式判断...")
    try:
        code_df = ak.stock_info_a_code_name()
        print("A股代码表前10条：")
        print(code_df.head(10))
        print("平安银行相关行：")
        print(code_df[code_df["code"].str.contains("000001")])
    except Exception as e:
        print(f"拉取A股代码表失败: {e}")

    stock_code = "sz000001"  # 平安银行
    start_date = "20150101"
    print("正在获取股票数据...")
    df = fetch_stock_data(stock_code, start_date)
    print("正在生成特征与标签...")
    processed = create_features_and_labels(df, n_past=5, n_future=3)
    # out_csv = os.path.join(os.path.dirname(__file__), f"{stock_code}_tabpfn.csv")
    out_csv =  f"{stock_code}_tabpfn.csv"
    print("正在保存为TabPFN格式...")
    save_for_tabpfn(processed, out_csv)
    print("预处理完成。")

if __name__ == "__main__":
    main()

预测

import os
# 临时绕过 /dev/null 权限问题
os.devnull = "/tmp/null"
if not os.path.exists("/tmp/null"):
    with open("/tmp/null", "w") as f:
        pass

import sys
import numpy as np
import pandas as pd

# 自动安装tabpfn（如未安装）
try:
    from tabpfn import TabPFNClassifier
except ImportError:
    import subprocess
    subprocess.check_call([sys.executable, "-m", "pip", "install", "tabpfn"])
    from tabpfn import TabPFNClassifier

# 读取数据
# DATA_PATH = os.path.join(os.path.dirname(__file__), "sz000001_tabpfn.csv")
DATA_PATH = "sz000001_tabpfn.csv"
data = pd.read_csv(DATA_PATH, header=None)
X = data.iloc[:, :-1].values
y = data.iloc[:, -1].values

# 按时间顺序划分（前80%训练，后20%测试）
split_idx = int(0.8 * len(X))
X_train, X_test = X[:split_idx], X[split_idx:]
y_train, y_test = y[:split_idx], y[split_idx:]

# TabPFN训练与预测
# clf = TabPFNClassifier(device='cpu')
clf = TabPFNClassifier(device='cuda')
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
accuracy = np.mean(y_pred == y_test)

print(f"Test Accuracy: {accuracy:.4f}")

最后生成的准确率: