前言
参加了Datawhale的AI夏令营第二期的机器学习赛道~，没错这次还是机器学习
baseline：https://aistudio.baidu.com/aistudio/projectdetail/6598302?sUid=2554132&shared=1&ts=1690895519028
赛事任务
简单地说就是通过模型预测股票价格
看这里： https://challenge.xfyun.cn/topic/info?type=quantitative-model
数据说明
- date：日期
- time：时间戳
- close：最新价/收盘价
- amount_delta：成交量变化 从上个tick到当前tick发生的成交金额
- n_midprice：中间价 标准化后的中间价，以涨跌幅表示
- n_bidN: 买N价
- n_bsizeN：买N量
- n_ask：卖N价
- n_asize1：卖N量
- labelN：Ntick价格移动方向 Ntick之后中间价相对于当前tick的移动方向，0为下跌，1为不变，2为上涨
评估指标
采用macro-F1 score进行评价，取label_5, label_10, label_20, label_40, label_60五项中的最高分作为最终得分。
F1 score解释： https://www.9998k.cn/archives/169.html
F1 score
F1 score = 2 * (precision * recall) / (precision + recall)
precision and recall
第一次看的时候还不太懂precision和recall的含义，也总结一下
首先定义以下几个概念：
TP（True Positive）：将本类归为本类
TN (True Negative) : 将其他类归为其他类
FP（False Positive）：错将其他类预测为本类
FN（False Negative）：本类标签预测为其他类标precision = TP / (TP + FP)
recall = TP / (TP + FN)
accuracy = (TP + TN) / (TP + TN + FP + FN)
分析
baseline分析
本次baseline使用了catboost，相比于上次的LightGBM，可以方便地调用显卡进行训练： 【Datawhale夏令营第二期】CatBoost如何使用GPU
在本次的baseline中，只需在cv_model函数中的line15中的params: dict中添加一个键即可：'task_type' : 'GPU'
代码解析
path = 'AI量化模型预测挑战赛公开数据/' # 数据目录
train_files = os.listdir(path+'train') # 获取目录下的文件
train_df = pd.DataFrame() # 定义一个空的DataFrame
for filename in tqdm.tqdm(train_files): # 遍历文件并使用tqdm显示
if os.path.isdir(path+'train/'+filename): # 先判断是否是缓存文件夹直接跳过，防止bug
continue
tmp = pd.read_csv(path+'train/'+filename) # 使用tmp变量存储读取转换后的DataFrame
tmp['file'] = filename # 记录文件名
train_df = pd.concat([train_df, tmp], axis=0, ignore_index=True) # 合并tmp数据到之前的df
test_files = os.listdir(path+'test')
test_df = pd.DataFrame()
for filename in tqdm.tqdm(test_files):
if os.path.isdir(path+'train/'+filename):
continue
tmp = pd.read_csv(path+'test/'+filename)
tmp['file'] = filename
test_df = pd.concat([test_df, tmp], axis=0, ignore_index=True)特征工程1.0
wap(Weighted Average Price)
加权平均价格 = (买价 买量 + 卖价 买量) / (买量 + 卖量)
# 计算wap数值
train_df['wap1'] = (train_df['n_bid1']*train_df['n_bsize1'] + train_df['n_ask1']*train_df['n_asize1'])/(train_df['n_bsize1'] + train_df['n_asize1'])
test_df['wap1'] = (test_df['n_bid1']*test_df['n_bsize1'] + test_df['n_ask1']*test_df['n_asize1'])/(test_df['n_bsize1'] + test_df['n_asize1'])
train_df['wap2'] = (train_df['n_bid2']*train_df['n_bsize2'] + train_df['n_ask2']*train_df['n_asize2'])/(train_df['n_bsize2'] + train_df['n_asize2'])
test_df['wap2'] = (test_df['n_bid2']*test_df['n_bsize2'] + test_df['n_ask2']*test_df['n_asize2'])/(test_df['n_bsize2'] + test_df['n_asize2'])
train_df['wap3'] = (train_df['n_bid3']*train_df['n_bsize3'] + train_df['n_ask3']*train_df['n_asize3'])/(train_df['n_bsize3'] + train_df['n_asize3'])
test_df['wap3'] = (test_df['n_bid3']*test_df['n_bsize3'] + test_df['n_ask3']*test_df['n_asize3'])/(test_df['n_bsize3'] + test_df['n_asize3'])
wap1
wap2
wap3
特征工程2.0
# 为了保证时间顺序的一致性，故进行排序
train_df = train_df.sort_values(['file','time'])
test_df = test_df.sort_values(['file','time'])
# 当前时间特征
# 围绕买卖价格和买卖量进行构建
# 暂时只构建买一卖一和买二卖二相关特征，进行优化时可以加上其余买卖信息
train_df['wap1'] = (train_df['n_bid1']*train_df['n_bsize1'] + train_df['n_ask1']*train_df['n_asize1'])/(train_df['n_bsize1'] + train_df['n_asize1'])
test_df['wap1'] = (test_df['n_bid1']*test_df['n_bsize1'] + test_df['n_ask1']*test_df['n_asize1'])/(test_df['n_bsize1'] + test_df['n_asize1'])
train_df['wap2'] = (train_df['n_bid2']*train_df['n_bsize2'] + train_df['n_ask2']*train_df['n_asize2'])/(train_df['n_bsize2'] + train_df['n_asize2'])
test_df['wap2'] = (test_df['n_bid2']*test_df['n_bsize2'] + test_df['n_ask2']*test_df['n_asize2'])/(test_df['n_bsize2'] + test_df['n_asize2'])
train_df['wap_balance'] = abs(train_df['wap1'] - train_df['wap2'])
train_df['price_spread'] = (train_df['n_ask1'] - train_df['n_bid1']) / ((train_df['n_ask1'] + train_df['n_bid1'])/2)
train_df['bid_spread'] = train_df['n_bid1'] - train_df['n_bid2']
train_df['ask_spread'] = train_df['n_ask1'] - train_df['n_ask2']
train_df['total_volume'] = (train_df['n_asize1'] + train_df['n_asize2']) + (train_df['n_bsize1'] + train_df['n_bsize2'])
train_df['volume_imbalance'] = abs((train_df['n_asize1'] + train_df['n_asize2']) - (train_df['n_bsize1'] + train_df['n_bsize2']))
test_df['wap_balance'] = abs(test_df['wap1'] - test_df['wap2'])
test_df['price_spread'] = (test_df['n_ask1'] - test_df['n_bid1']) / ((test_df['n_ask1'] + test_df['n_bid1'])/2)
test_df['bid_spread'] = test_df['n_bid1'] - test_df['n_bid2']
test_df['ask_spread'] = test_df['n_ask1'] - test_df['n_ask2']
test_df['total_volume'] = (test_df['n_asize1'] + test_df['n_asize2']) + (test_df['n_bsize1'] + test_df['n_bsize2'])
test_df['volume_imbalance'] = abs((test_df['n_asize1'] + test_df['n_asize2']) - (test_df['n_bsize1'] + test_df['n_bsize2']))
# 历史平移
# 获取历史信息
for val in ['wap1','wap2','wap_balance','price_spread','bid_spread','ask_spread','total_volume','volume_imbalance']:
for loc in [1,5,10,20,40,60]:
train_df[f'file_{val}_shift{loc}'] = train_df.groupby(['file'])[val].shift(loc)
test_df[f'file_{val}_shift{loc}'] = test_df.groupby(['file'])[val].shift(loc)
# 差分特征
# 获取与历史数据的增长关系
for val in ['wap1','wap2','wap_balance','price_spread','bid_spread','ask_spread','total_volume','volume_imbalance']:
for loc in [1,5,10,20,40,60]:
train_df[f'file_{val}_diff{loc}'] = train_df.groupby(['file'])[val].diff(loc)
test_df[f'file_{val}_diff{loc}'] = test_df.groupby(['file'])[val].diff(loc)
# 窗口统计
# 获取历史信息分布变化信息
# 可以尝试更多窗口大小已经统计方式，如min、max、median等
for val in ['wap1','wap2','wap_balance','price_spread','bid_spread','ask_spread','total_volume','volume_imbalance']:
train_df[f'file_{val}_win7_mean'] = train_df.groupby(['file'])[val].transform(lambda x: x.rolling(window=7, min_periods=3).mean())
train_df[f'file_{val}_win7_std'] = train_df.groupby(['file'])[val].transform(lambda x: x.rolling(window=7, min_periods=3).std())
test_df[f'file_{val}_win7_mean'] = test_df.groupby(['file'])[val].transform(lambda x: x.rolling(window=7, min_periods=3).mean())
test_df[f'file_{val}_win7_std'] = test_df.groupby(['file'])[val].transform(lambda x: x.rolling(window=7, min_periods=3).std())
# 时间相关特征
train_df['hour'] = train_df['time'].apply(lambda x:int(x.split(':')[0]))
test_df['hour'] = test_df['time'].apply(lambda x:int(x.split(':')[0]))
train_df['minute'] = train_df['time'].apply(lambda x:int(x.split(':')[1]))
test_df['minute'] = test_df['time'].apply(lambda x:int(x.split(':')[1]))
# 入模特征
cols = [f for f in test_df.columns if f not in ['uuid','time','file']]模型训练
def cv_model(clf, train_x, train_y, test_x, clf_name, seed = 2023):
folds = 5
kf = KFold(n_splits=folds, shuffle=True, random_state=seed)
oof = np.zeros([train_x.shape[0], 3])
test_predict = np.zeros([test_x.shape[0], 3])
cv_scores = []
for i, (train_index, valid_index) in enumerate(kf.split(train_x, train_y)):
print('************************************ {} ************************************'.format(str(i+1)))
trn_x, trn_y, val_x, val_y = train_x.iloc[train_index], train_y[train_index], train_x.iloc[valid_index], train_y[valid_index]
if clf_name == "cat":
params = {'learning_rate': 0.12, 'depth': 9, 'bootstrap_type':'Bernoulli','random_seed':2023,
'od_type': 'Iter', 'od_wait': 100, 'random_seed': 11, 'allow_writing_files': False,
'loss_function': 'MultiClass', 'task_type' : 'GPU'} # 使用'task_type' : 'GPU'可利用gpu加速训练
model = clf(iterations=2000, **params)
model.fit(trn_x, trn_y, eval_set=(val_x, val_y),
metric_period=200,
use_best_model=True,
cat_features=[],
verbose=1)
val_pred = model.predict_proba(val_x)
test_pred = model.predict_proba(test_x)
oof[valid_index] = val_pred
test_predict += test_pred / kf.n_splits
F1_score = f1_score(val_y, np.argmax(val_pred, axis=1), average='macro')
cv_scores.append(F1_score)
print(cv_scores)
return oof, test_predict
for label in ['label_5','label_10','label_20','label_40','label_60']:
print(f'=================== {label} ===================')
cat_oof, cat_test = cv_model(CatBoostClassifier, train_df[cols], train_df[label], test_df[cols], 'cat')
train_df[label] = np.argmax(cat_oof, axis=1)
test_df[label] = np.argmax(cat_test, axis=1)疑问
为什么没有使用rnn或lstm进行预测？
评论 (0)