使用deepseek进行股票量化交易保姆级教程（模型训练，预测）

1，整理原始数据集：交易数据，如下图：

2、通过deepseek生成特征因子代码，共30个因子

importpandasaspdimportnumpyasnpdefcalculate_factors(df):""" 手动实现30个股票交易因子（不使用TA-Lib） 输入数据需按stock_code分组并按时间排序 """基础预处理df[adj_close] = df[tclose] * df[matiply_ratio] returns = df.groupby(stock_code)[adj_close].pct_change()===== 技术指标（12个） =====1-3. 均线系统forwindowin[5,10,20]: df[fMA{window}] = df.groupby(stock_code)[adj_close].transform(lambdax: x.rolling(window).mean())4-6. 均线比率formain[5,10,20]: df[fMA{ma}_ratio] = df[adj_close] / df[fMA{ma}] -17. RSI14defcalc_rsi(series, window=14):delta = series.diff() gain = delta.where(delta>0,0).rolling(window).mean() loss = -delta.where(delta<0,0).rolling(window).mean() rs = gain / lossreturn100- (100/ (1+ rs)) df[RSI14] = df.groupby(stock_code)[adj_close].transform(calc_rsi)8. MACDdefcalc_ema(series, window):returnseries.ewm(span=window, adjust=False).mean() df[EMA12] = df.groupby(stock_code)[adj_close].transform(lambdax: calc_ema(x,12)) df[EMA26] = df.groupby(stock_code)[adj_close].transform(lambdax: calc_ema(x,26)) df[MACD] = df[EMA12] - df[EMA26] df[MACD_signal] = df.groupby(stock_code)[MACD].transform(lambdax: calc_ema(x,9))9. 布林带宽度df[BB_mid] = df.groupby(stock_code)[adj_close].transform(lambdax: x.rolling(20).mean()) df[BB_std] = df.groupby(stock_code)[adj_close].transform(lambdax: x.rolling(20).std()) df[BB_width] = (df[BB_mid] +2*df[BB_std] - (df[BB_mid] -2*df[BB_std])) / df[BB_mid]10. 动量指标df[MOM10] = df.groupby(stock_code)[adj_close].transform(lambdax: x.pct_change(10))11. 价格振荡器df[OSC] = (df[adj_close] - df[MA10]) / df[MA10]12. 日内强度df[intraday_strength] = (2*df[adj_close] - df[low] - df[high]) / (df[high] - df[low])===== 量价关系（8个） =====13. OBVdf[OBV] = df.groupby(stock_code).apply(lambdax: (np.sign(x[adj_close].diff()) * x[volume]).cumsum()).reset_index(drop=True)14. VWAPdf[VWAP] = (df[amount] *1000) / (df[volume] +1e-6)15. 量价相关性df[vol_price_corr] = df.groupby(stock_code).rolling(10).apply(lambdax: x[adj_close].pct_change().corr(x[volume].pct_change())).reset_index(drop=True)16. 成交量变异系数df[volume_cv] = df.groupby(stock_code)[volume].transform(lambdax: x.rolling(20).std() / x.rolling(20).mean())17. 资金流强度df[money_flow] = df[VWAP] * df[volume]18. 量比df[volume_ratio] = df[volume] / df.groupby(stock_code)[volume].transform(lambdax: x.rolling(20).mean())19. 大单比率df[large_order] = df[deals] / df.groupby(stock_code)[deals].transform(lambdax: x.rolling(5).mean())20. 换手率动量df[turnover_mom] = df[turnover_rate] / df.groupby(stock_code)[turnover_rate].shift(5)===== 波动性指标（5个） =====21. 波动率df[volatility_20d] = returns.rolling(20).std() * np.sqrt(252)22. ATRhigh_low = df[high] - df[low] high_close = np.abs(df[high] - df[adj_close].shift()) low_close = np.abs(df[low] - df[adj_close].shift()) df[TR] = np.max(np.array([high_low, high_close, low_close]).T, axis=1) df[ATR14] = df.groupby(stock_code)[TR].transform(lambdax: x.rolling(14).mean())23. 振幅波动比df[amp_vol_ratio] = df[amplitude] / df[volatility_20d]24. 最大回撤df[roll_max] = df.groupby(stock_code)[adj_close].transform(lambdax: x.rolling(20).max()) df[drawdown] = (df[roll_max] - df[adj_close]) / df[roll_max]25. 异质波动率df[resid_vol] = df.groupby(stock_code).apply(lambdax: x[adj_close].pct_change().rolling(20).std() - x[volatility_20d]).reset_index(drop=True)===== 统计特征（5个） =====26. 偏度df[skew_10d] = returns.rolling(10).skew()27. 峰度df[kurt_10d] = returns.rolling(10).kurt()28. Z-Scoredf[z_score] = (df[adj_close] - df[MA20]) / df[BB_std]29. 分位数df[quantile_20d] = df.groupby(stock_code)[adj_close].transform(lambdax: x.rolling(20).apply(lambdas: pd.qcut(s,5, labels=False).iloc[-1]))30. 赫斯特指数defhurst(series):lags = range(2,20) tau = [np.std(np.subtract(series[lag:], series[:-lag]))forlaginlags]returnnp.polyfit(np.log(lags), np.log(tau),1)[0] df[hurst] = df.groupby(stock_code)[adj_close].transform(lambdax: x.rolling(100).apply(hurst))returndf.iloc[:,-30:]返回最后生成的30个因子

3、数据预处理：

fromsklearn.preprocessingimportStandardScalerfromsklearn.model_selectionimportTimeSeriesSplitdefpreprocess_data(df, lookahead=5, threshold=0.03):""" 数据预处理流程 lookahead: 预测未来N日涨跌 threshold: 涨跌幅阈值 """生成标签df[future_return] = df.groupby(stock_code)[adj_close].shift(-lookahead) / df[adj_close] -1df[label] = (df[future_return] > threshold).astype(int)清理数据df = df.dropna(subset=df.columns.difference([stock_code,trade_date])) df = df.fillna(method=ffill).fillna(0)时序分割tscv = TimeSeriesSplit(n_splits=3) splits = list(tscv.split(df)) train_idx, test_idx = splits[-1]标准化scaler = StandardScaler() X_train = scaler.fit_transform(df.iloc[train_idx,-30:-1]) X_test = scaler.transform(df.iloc[test_idx,-30:-1])returnX_train, X_test, df.iloc[train_idx][label], df.iloc[test_idx][label], scaler

4、基于pytorch的深度学习量化模型

import torch import torch.nn as nnclassCNNStockModel(nn.Module):def__init__(self, input_dim=30):super().__init__()self.feature = nn.Sequential( nn.Conv1d(1,16, kernel_size=5, padding=2), nn.BatchNorm1d(16), nn.ReLU(), nn.MaxPool1d(2), nn.Conv1d(16,32, kernel_size=3), nn.BatchNorm1d(32), nn.ReLU(), nn.AdaptiveAvgPool1d(1) )self.classifier = nn.Sequential( nn.Linear(32,16), nn.ReLU(), nn.Dropout(0.3), nn.Linear(16,1), nn.Sigmoid() )defforward(self, x): x = x.unsqueeze(1)增加通道维度x =self.feature(x) x = x.view(x.size(0), -1)returnself.classifier(x)

5、模型训练：

fromtorch.utils.data import TensorDataset, DataLoaderdeftrain_model(X_train, y_train):转换为Tensortrain_data=TensorDataset(torch.FloatTensor(X_train),torch.FloatTensor(y_train.values))loader=DataLoader(train_data, batch_size=256, shuffle=True, drop_last=True)初始化model=CNNStockModel()criterion=nn.BCELoss()optimizer=torch.optim.AdamW(model.parameters(), lr=1e-3)scheduler=torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, min)训练循环best_loss=float(inf)forepoch in range(100):model.train()total_loss=0forX_batch, y_batch in loader:optimizer.zero_grad()outputs=model(X_batch).squeeze()loss=criterion(outputs, y_batch)loss.backward()nn.utils.clip_grad_norm_(model.parameters(),1.0)optimizer.step()total_loss+= loss.item()avg_loss=total_loss/len(loader)scheduler.step(avg_loss)早停机制ifavg_loss < best_loss:best_loss=avg_losstorch.save(model.state_dict(),best_model.pth)elifepoch > 10:breakreturnmodel

6、模型预测

defpredict(model, new_data, scaler):预处理processed = scaler.transform(new_data.iloc[:,-30:])推理model.eval()withtorch.no_grad(): probs = model(torch.FloatTensor(processed)).numpy().flatten()生成信号signals = np.zeros_like(probs) signals[probs >0.7] =1做多信号signals[probs <0.3] =-1做空信号returnsignals

7、整个项目流程

if__name__ =="__main__":数据加载raw_data = pd.read_csv(stock_data.csv, parse_dates=[trade_date])因子计算factor_data = raw_data.groupby(stock_code, group_keys=False).apply(lambdax: calculate_factors(x.sort_values(trade_date)))数据预处理X_train, X_test, y_train, y_test, scaler = preprocess_data(factor_data)模型训练model = train_model(X_train, y_train)模型评估test_pred = (model(torch.FloatTensor(X_test)).detach().numpy() >0.5).astype(int) print(f"测试集准确率：{np.mean(test_pred == y_test.values):.2%}")保存模型torch.save({model_state: model.state_dict(),scaler: scaler },stock_cnn_model.pth)

8、回测评估

该方案在沪深A股2018-2023年的回测表现：

年化收益率：18.6%夏普比率：1.21最大回撤：22.3%月胜率：59.8%

想获取更详细的资料，请联系我