作者：flaty

日本炒股大赛最稳定的金牌方案 

简介

第四名的选手是从第一轮到最后一轮都一直稳居在前排的选手，所以其方案在日本炒股大赛中是非常稳定的，其背后究竟有何秘密，我们一起来学习一下它的开源代码。

方案

01

数据预处理

• 对收盘价进行价格的调整

def adjust_price(price):
    """
    Args:
        price (pd.DataFrame)  : pd.DataFrame include stock_price
    Returns:
        price DataFrame (pd.DataFrame): stock_price with generated AdjustedClose
    """
    # transform Date column into datetime
    price.loc[: ,"Date"] = pd.to_datetime(price.loc[: ,"Date"], format="%Y-%m-%d")

    def generate_adjusted_close(df):
        """
        Args:
            df (pd.DataFrame)  : stock_price for a single SecuritiesCode
        Returns:
            df (pd.DataFrame): stock_price with AdjustedClose for a single SecuritiesCode
        """
        # sort data to generate CumulativeAdjustmentFactor
        df = df.sort_values("Date", ascending=False)#降順(最新のものが先頭)
        # generate CumulativeAdjustmentFactor
        df.loc[:, "CumulativeAdjustmentFactor"] = df["AdjustmentFactor"].cumprod()#cumprodは累積積を求める関数
        # generate AdjustedClose
        df.loc[:, "AdjustedClose"] = (
            df["CumulativeAdjustmentFactor"] * df["Close"]
        ).map(lambda x: float(
            Decimal(str(x)).quantize(Decimal('0.1'), rounding=ROUND_HALF_UP)#四捨五入
        ))
        # reverse order
        df = df.sort_values("Date")#昇順に戻す
        # to fill AdjustedClose, replace 0 into np.nan
        df.loc[df["AdjustedClose"] == 0, "AdjustedClose"] = np.nan
        # forward fill AdjustedClose
        df.loc[:, "AdjustedClose"] = df.loc[:, "AdjustedClose"].ffill()#ffill:前(上)の値に置換
        return df

    # generate AdjustedClose
    price = price.sort_values(["SecuritiesCode", "Date"])
    price = price.groupby("SecuritiesCode").apply(generate_adjusted_close).reset_index(drop=True)

    price.set_index("Date", inplace=True)
    return price

02

特征

• 一天的收益回报率

• ExpectedDividend

def get_features_for_predict(price, code):
    """
    Args:
        price (pd.DataFrame)  : pd.DataFrame include stock_price
        code (int)  : A local code for a listed company
    Returns:
        feature DataFrame (pd.DataFrame)
    """
    close_col = "AdjustedClose"
    feats = price.loc[price["SecuritiesCode"] == code, ["SecuritiesCode", close_col, "ExpectedDividend"]].copy()

    # calculate return using AdjustedClose
    feats["return_1day"] = feats[close_col].pct_change(1)
    
    # ExpectedDividend
    feats["ExpectedDividend"] = feats["ExpectedDividend"].mask(feats["ExpectedDividend"] > 0, 1)

    # filling data for nan and inf
    feats = feats.fillna(0)
    feats = feats.replace([np.inf, -np.inf], 0)
    # drop AdjustedClose column
    feats = feats.drop([close_col], axis=1)

    return feats

03

模型预测

作者没有构建复杂的模型，只是两个特征进行简单的想加，这个是非常值得学习和思考的。

df_price = adjust_price(df_price_raw) 
    # get target SecuritiesCodes
    codes = sorted(prices["SecuritiesCode"].unique())

    # generate feature
    feature = pd.concat([get_features_for_predict(df_price, code) for code in codes])
    # filter feature for this iteration
    feature = feature.loc[feature.index == current_date]

    # prediction
    feature.loc[:, "predict"] = feature["return_1day"] + feature["ExpectedDividend"]*100

    # set rank by predict
    feature = feature.sort_values("predict", ascending=True).drop_duplicates(subset=['SecuritiesCode'])
    feature.loc[:, "Rank"] = np.arange(len(feature))
    feature_map = feature.set_index('SecuritiesCode')['Rank'].to_dict()
    sample_prediction['Rank'] = sample_prediction['SecuritiesCode'].map(feature_map)

参考文献

1. https://www.kaggle.com/code/flat831/4th-place-model/notebook?scriptVersionId=100052889

2. https://www.kaggle.com/competitions/jpx-tokyo-stock-exchange-prediction/discussion/359151

往期精彩回顾




适合初学者入门人工智能的路线及资料下载(图文+视频)机器学习入门系列下载机器学习及深度学习笔记等资料打印《统计学习方法》的代码复现专辑机器学习交流qq群955171419，加入微信群请扫码