金融数据_PySpark-3.0.3梯度提升树(GBTClassifier)实例

Mostcow2024-04-10 15:54

金融数据_PySpark-3.0.3梯度提升树(GBTClassifier)实例

梯度提升树 (Gradient Boosting Tree): 梯度提升树是一种集成学习方法, 可以通过组合多个弱学习器提高模型性能。

对于分类问题, 可以使用梯度提升决策树。

在实际应用中, 你可能需要进行一些特征工程, 确保输入特征的质量和多样性。

此外, 股票市场的预测问题本身非常复杂, 因为受到众多因素的影响, 包括市场情绪、宏观经济指标等。

过去的股票价格并不一定能够准确预测未来的价格, 因此在实际应用中, 你需要考虑到这些限制。

最好的模型可能会根据具体的数据和问题而异, 因此建议尝试多个模型, 进行交叉验证, 并评估它们在你的数据集上的性能。

txt 复制代码 
* 当使用 PySpark 进行梯度提升树的构建时, 你可以使用 GBTClassifier 类。

* 下面是一个简单的示例, 演示如何在 PySpark 中构建和训练梯度提升树分类器。

实例数据

本实例截取了 "湖北宜化(000422)" 2015年08月06日 - 2015年12月31日的数据。

HBYH_000422_20150806_20151231.csv

csv 复制代码 
Date,Code,Open,High,Low,Close,Pre_Close,Change,Turnover_Rate,Volume,MA5,MA10
2015-12-31,'000422,7.93,7.95,7.76,7.77,7.93,-0.020177,0.015498,13915200,7.86,7.85
2015-12-30,'000422,7.86,7.93,7.75,7.93,7.84,0.011480,0.018662,16755900,7.90,7.85
2015-12-29,'000422,7.72,7.85,7.69,7.84,7.71,0.016861,0.015886,14263800,7.90,7.81
2015-12-28,'000422,8.03,8.08,7.70,7.71,8.03,-0.039851,0.030821,27672800,7.91,7.78
2015-12-25,'000422,8.03,8.05,7.93,8.03,7.99,0.005006,0.021132,18974000,7.93,7.78
2015-12-24,'000422,7.93,8.16,7.87,7.99,7.92,0.008838,0.026487,23781900,7.85,7.72
2015-12-23,'000422,7.97,8.11,7.88,7.92,7.89,0.003802,0.042360,38033600,7.80,7.69
2015-12-22,'000422,7.86,7.93,7.76,7.89,7.83,0.007663,0.026929,24178700,7.73,7.68
2015-12-21,'000422,7.59,7.89,7.56,7.83,7.63,0.026212,0.030777,27633600,7.66,7.67
2015-12-18,'000422,7.71,7.74,7.57,7.63,7.74,-0.014212,0.024764,22234900,7.62,7.71
2015-12-17,'000422,7.58,7.75,7.57,7.74,7.55,0.025166,0.028054,25188400,7.59,7.77
2015-12-16,'000422,7.57,7.62,7.53,7.55,7.55,0.000000,0.020718,18601600,7.58,7.79
2015-12-15,'000422,7.63,7.66,7.52,7.55,7.62,-0.009186,0.025902,23256600,7.64,7.78
2015-12-14,'000422,7.40,7.64,7.36,7.62,7.51,0.014647,0.021005,18860100,7.68,7.76
2015-12-11,'000422,7.65,7.70,7.41,7.51,7.67,-0.020860,0.020477,18385900,7.80,7.73
2015-12-10,'000422,7.78,7.87,7.65,7.67,7.83,-0.020434,0.019972,17931900,7.95,7.69
2015-12-09,'000422,7.76,8.00,7.75,7.83,7.77,0.007722,0.025137,22569700,8.00,7.68
2015-12-08,'000422,8.08,8.18,7.76,7.77,8.24,-0.057039,0.036696,32948200,7.92,7.66
2015-12-07,'000422,8.12,8.39,7.94,8.24,8.23,0.001215,0.064590,57993100,7.84,7.64
2015-12-04,'000422,7.85,8.48,7.80,8.23,7.92,0.039141,0.100106,89881900,7.65,7.58
2015-12-03,'000422,7.42,8.09,7.38,7.92,7.43,0.065949,0.045416,40777500,7.43,7.52
2015-12-02,'000422,7.35,7.48,7.20,7.43,7.36,0.009511,0.015968,14337600,7.37,7.49
2015-12-01,'000422,7.28,7.39,7.23,7.36,7.33,0.004093,0.012308,11050700,7.41,7.48
2015-11-30,'000422,7.18,7.36,6.95,7.33,7.11,0.030942,0.020323,18247500,7.45,7.50
2015-11-27,'000422,7.59,7.59,6.95,7.11,7.60,-0.064474,0.027673,24846700,7.51,7.52
2015-11-26,'000422,7.63,7.73,7.58,7.60,7.63,-0.003932,0.024836,22299800,7.61,7.54
2015-11-25,'000422,7.56,7.64,7.51,7.63,7.59,0.005270,0.020919,18782900,7.61,7.54
2015-11-24,'000422,7.60,7.63,7.48,7.59,7.62,-0.003937,0.014867,13348200,7.56,7.53
2015-11-23,'000422,7.59,7.72,7.55,7.62,7.61,0.001314,0.028406,25505000,7.54,7.53
2015-11-20,'000422,7.59,7.71,7.53,7.61,7.59,0.002635,0.028277,25389100,7.52,7.53
2015-11-19,'000422,7.45,7.62,7.41,7.59,7.39,0.027064,0.038638,34691700,7.47,7.52
2015-11-18,'000422,7.53,7.54,7.38,7.39,7.51,-0.015979,0.014173,12725000,7.46,7.50
2015-11-17,'000422,7.53,7.63,7.44,7.51,7.50,0.001333,0.028640,25714500,7.51,7.50
2015-11-16,'000422,7.27,7.52,7.24,7.50,7.38,0.016260,0.016230,14572000,7.52,7.46
2015-11-13,'000422,7.49,7.55,7.36,7.38,7.54,-0.021220,0.029196,26214400,7.53,7.41
2015-11-12,'000422,7.65,7.68,7.49,7.54,7.61,-0.009198,0.026501,23794800,7.56,7.40
2015-11-11,'000422,7.57,7.64,7.52,7.61,7.57,0.005284,0.026113,23445900,7.54,7.37
2015-11-10,'000422,7.51,7.61,7.45,7.57,7.55,0.002649,0.024979,22427700,7.49,7.32
2015-11-09,'000422,7.51,7.62,7.45,7.55,7.53,0.002656,0.033367,29959500,7.39,7.31
2015-11-06,'000422,7.47,7.53,7.37,7.53,7.45,0.010738,0.037058,33273100,7.29,7.27
2015-11-05,'000422,7.34,7.54,7.32,7.45,7.37,0.010855,0.040463,36330200,7.24,7.24
2015-11-04,'000422,7.10,7.38,7.07,7.37,7.05,0.045390,0.034817,31260800,7.20,7.17
2015-11-03,'000422,7.08,7.13,7.02,7.05,7.06,-0.001416,0.014938,13412400,7.15,7.10
2015-11-02,'000422,7.11,7.26,7.05,7.06,7.26,-0.027548,0.016865,15142100,7.23,7.10
2015-10-30,'000422,7.22,7.38,7.10,7.26,7.24,0.002762,0.022821,20490200,7.25,7.10
2015-10-29,'000422,7.27,7.33,7.16,7.24,7.16,0.011173,0.025726,23098500,7.23,7.08
2015-10-28,'000422,7.32,7.40,7.09,7.16,7.42,-0.035040,0.035572,31938500,7.15,7.05
2015-10-27,'000422,7.21,7.48,7.08,7.42,7.18,0.033426,0.057658,51769300,7.04,7.01
2015-10-26,'000422,7.20,7.25,7.01,7.18,7.17,0.001395,0.036840,33077800,6.98,6.96
2015-10-23,'000422,6.84,7.22,6.81,7.17,6.80,0.054412,0.047169,42351500,6.95,6.93
2015-10-22,'000422,6.68,6.81,6.64,6.80,6.65,0.022556,0.020609,18503800,6.93,6.87
2015-10-21,'000422,7.08,7.11,6.61,6.65,7.09,-0.062059,0.039388,35365300,6.96,6.85
2015-10-20,'000422,7.00,7.09,6.94,7.09,7.03,0.008535,0.024472,21972900,6.98,6.81
2015-10-19,'000422,7.09,7.13,6.92,7.03,7.08,-0.007062,0.031262,28068800,6.94,6.72
2015-10-16,'000422,6.97,7.08,6.91,7.08,6.93,0.021645,0.039632,35584700,6.91,6.66
2015-10-15,'000422,6.77,6.94,6.75,6.93,6.77,0.023634,0.031645,28412700,6.82,6.59
2015-10-14,'000422,6.87,6.94,6.74,6.77,6.89,-0.017417,0.027226,24445500,6.74,6.55
2015-10-13,'000422,6.86,6.96,6.80,6.89,6.88,0.001453,0.028704,25771900,6.64,6.51
2015-10-12,'000422,6.62,6.91,6.58,6.88,6.61,0.040847,0.037037,33254300,6.50,6.49
2015-10-09,'000422,6.54,6.65,6.45,6.61,6.54,0.010703,0.018528,16635900,6.41,6.46
2015-10-08,'000422,6.45,6.70,6.37,6.54,6.26,0.044728,0.018857,16931000,6.35,6.44
2015-09-30,'000422,6.25,6.30,6.22,6.26,6.23,0.004815,0.007327,6579090,6.35,6.43
2015-09-29,'000422,6.30,6.32,6.18,6.23,6.40,-0.026562,0.008991,8072900,6.39,6.48
2015-09-28,'000422,6.35,6.42,6.25,6.40,6.34,0.009464,0.008824,7922890,6.48,6.47
2015-09-25,'000422,6.51,6.56,6.25,6.34,6.53,-0.029096,0.012584,11298800,6.51,6.45
2015-09-24,'000422,6.48,6.56,6.45,6.53,6.45,0.012403,0.011339,10180900,6.53,6.51
2015-09-23,'000422,6.51,6.60,6.41,6.45,6.67,-0.032984,0.015920,14294100,6.52,6.54
2015-09-22,'000422,6.58,6.73,6.54,6.67,6.58,0.013678,0.023356,20970200,6.56,6.60
2015-09-21,'000422,6.34,6.61,6.29,6.58,6.44,0.021739,0.017036,15295900,6.46,6.62
2015-09-18,'000422,6.52,6.58,6.30,6.44,6.44,0.000000,0.016622,14924700,6.39,6.62
2015-09-17,'000422,6.59,6.76,6.43,6.44,6.68,-0.035928,0.019517,17523900,6.48,6.62
2015-09-16,'000422,6.21,6.76,6.17,6.68,6.15,0.086179,0.019671,17662300,6.56,6.65
2015-09-15,'000422,6.24,6.38,6.05,6.15,6.26,-0.017572,0.015338,13771200,6.64,6.66
2015-09-14,'000422,6.89,6.95,6.18,6.26,6.87,-0.088792,0.021233,18559600,6.78,6.75
2015-09-11,'000422,6.87,6.96,6.77,6.87,6.84,0.004386,0.010853,9486290,6.85,6.79
2015-09-10,'000422,6.95,7.01,6.76,6.84,7.06,-0.031161,0.017423,15229100,6.76,6.74
2015-09-09,'000422,6.90,7.09,6.86,7.06,6.88,0.026163,0.028974,25325600,6.74,6.68
2015-09-08,'000422,6.65,6.91,6.55,6.88,6.62,0.039275,0.017858,15609100,6.69,6.67
2015-09-07,'000422,6.50,6.81,6.50,6.62,6.38,0.037618,0.017850,15602600,6.72,6.75
2015-09-02,'000422,6.45,6.88,6.30,6.38,6.74,-0.053412,0.022286,19480100,6.73,6.91
2015-09-01,'000422,6.88,6.99,6.67,6.74,6.81,-0.010279,0.025829,22576700,6.72,7.12
2015-08-31,'000422,6.90,6.97,6.71,6.81,7.07,-0.036775,0.018385,16069600,6.62,7.24
2015-08-28,'000422,6.75,7.08,6.71,7.07,6.67,0.059970,0.026692,23330800,6.65,7.44
2015-08-27,'000422,6.53,6.67,6.34,6.67,6.32,0.055380,0.022455,19627900,6.78,7.59
2015-08-26,'000422,6.31,6.77,6.09,6.32,6.25,0.011200,0.029963,26190200,7.08,7.76
2015-08-25,'000422,6.40,6.77,6.25,6.25,6.94,-0.099424,0.029492,25778600,7.52,7.96
2015-08-24,'000422,7.49,7.49,6.94,6.94,7.71,-0.099870,0.036552,31949900,7.86,8.18
2015-08-21,'000422,8.00,8.11,7.60,7.71,8.17,-0.056304,0.032199,28144800,8.23,8.33
2015-08-20,'000422,8.38,8.56,8.14,8.17,8.53,-0.042204,0.031764,27764200,8.40,8.38
2015-08-19,'000422,7.73,8.57,7.72,8.53,7.96,0.071608,0.052192,45619900,8.45,8.37
2015-08-18,'000422,8.81,8.86,7.92,7.96,8.80,-0.095455,0.056179,49105500,8.39,8.32
2015-08-17,'000422,8.49,8.83,8.42,8.80,8.52,0.032864,0.048161,42096900,8.50,8.35
2015-08-14,'000422,8.48,8.65,8.43,8.52,8.44,0.009479,0.041169,35985000,8.43,8.24
2015-08-13,'000422,8.20,8.45,8.15,8.44,8.24,0.024272,0.029768,26019600,8.37,8.16
2015-08-12,'000422,8.38,8.48,8.21,8.24,8.48,-0.028302,0.035421,30960700,8.30,8.08
2015-08-11,'000422,8.41,8.68,8.32,8.48,8.49,-0.001178,0.048444,42343900,8.26,8.03
2015-08-10,'000422,8.28,8.58,8.18,8.49,8.21,0.034105,0.041268,36071600,8.20,7.92
2015-08-07,'000422,8.15,8.28,8.08,8.21,8.07,0.017348,0.025855,22599800,8.05,7.81
2015-08-06,'000422,7.88,8.21,7.80,8.07,8.03,0.004981,0.020074,17546700,7.95,7.80

探索思路

这里只是简单示例, 目的在于熟悉 Spark 中的梯度提升树分类器使用方法, 无任何投资引导。

目标:

通过当日数值情况, 预测当日收盘涨跌, 如果 "涨跌幅(Change) >= 0", 则用 1 表示, 如果 "涨跌幅(Change) < 0", 则用 0 表示 (二分类标签)。

变量:

  1. 当日最高价

  2. 当日最低价

  3. 当日换手率

  4. 当日成交量

  5. 当日星期几 (星期对价格的影响)

  6. 当日 "短期均线(MA5)" 与 "长期均线(MA10)" 的关系, 如果 "MA5 > MA10", 则用 1 表示, 如果 "MA5 = MA10", 则用 0 表示, 如果 "MA5 < MA10", 则用 -1 表示。

导入 pyspark.sql 相关模块

Spark SQL 是用于结构化数据处理的 Spark 模块。它提供了一种成为 DataFrame 编程抽象, 是由 SchemaRDD 发展而来。

不同于 SchemaRDD 直接继承 RDD, DataFrame 自己实现了 RDD 的绝大多数功能。

python 复制代码 
from pyspark.sql import Row, SparkSession
from pyspark.sql.functions import col
from pyspark.sql.types import DateType, IntegerType, DoubleType

导入 pyspark.ml 相关模块

Spark 在核心数据抽象 RDD 的基础上, 支持 4 大组件, 其中机器学习占其一。

进一步的, Spark 中实际上支持两个机器学习模块, MLlib 和 ML, 区别在于前者主要是基于 RDD 数据结构, 当前处于维护状态; 而后者则是 DataFrame 数据结构, 支持更多的算法, 后续将以此为主进行迭代。

所以, 在实际应用中优先使用 ML 子模块。

Spark 的 ML 库与 Python 中的另一大机器学习库 Sklearn 的关系是: Spark 的 ML 库支持大部分机器学习算法和接口功能, 虽远不如 Sklearn 功能全面, 但主要面向分布式训练, 针对大数据。

而 Sklearn 是单点机器学习算法库, 支持几乎所有主流的机器学习算法, 从样例数据, 特征选择, 模型选择和验证, 基础学习算法和集成学习算法, 提供了机器学习一站式解决方案, 但仅支持并行而不支持分布式。

python 复制代码 
from pyspark.ml.feature import StringIndexer, VectorAssembler
from pyspark.ml.classification import GBTClassifier
from pyspark.ml.evaluation import BinaryClassificationEvaluator
from pyspark.ml import Pipeline

创建 SparkSession 对象

Spark 2.0 以上版本的 spark-shell 在启动时会自动创建一个名为 spark 的 SparkSession 对象。

当需要手工创建时, SparkSession 可以由其伴生对象的 builder 方法创建出来。

python 复制代码 
spark = SparkSession.builder.master("local[*]").appName("spark").getOrCreate()

使用 Spark 构建 DataFrame 数据 (Optional)

当数据量较小时, 可以使用该方法手工构建 DataFrame 数据。

构建数据行 Row (以前 3 行为例):

python 复制代码 
Row(Date="2015-12-31", Code="'000422", Open="7.93", High="7.95", Low="7.76", Close="7.77", Pre_Close="7.93", Change="-0.020177", Turnover_Rate="0.015498", Volume="13915200", MA5="7.86", MA10="7.85")
ROW(Date="2015-12-30", Code="'000422", Open="7.86", High="7.93", Low="7.75", Close="7.93", Pre_Close="7.84", Change="0.011480", Turnover_Rate="0.018662", Volume="16755900", MA5="7.90", MA10="7.85")
Row(Date="2015-12-29", Code="'000422", Open="7.72", High="7.85", Low="7.69", Close="7.84", Pre_Close="7.71", Change="0.016861", Turnover_Rate="0.015886", Volume="14263800", MA5="7.90", MA10="7.81")

将构建好的数据行 Row 加入列表 (以前 3 行为例):

python 复制代码 
Data_Rows = [
    Row(Date="2015-12-31", Code="'000422", Open="7.93", High="7.95", Low="7.76", Close="7.77", Pre_Close="7.93", Change="-0.020177", Turnover_Rate="0.015498", Volume="13915200", MA5="7.86", MA10="7.85"),
    ROW(Date="2015-12-30", Code="'000422", Open="7.86", High="7.93", Low="7.75", Close="7.93", Pre_Close="7.84", Change="0.011480", Turnover_Rate="0.018662", Volume="16755900", MA5="7.90", MA10="7.85"),
    Row(Date="2015-12-29", Code="'000422", Open="7.72", High="7.85", Low="7.69", Close="7.84", Pre_Close="7.71", Change="0.016861", Turnover_Rate="0.015886", Volume="14263800", MA5="7.90", MA10="7.81")
]

生成 DataFrame 数据框 (以前 3 行为例):

python 复制代码 
SDF = spark.createDataFrame(Data_Rows)

输出 DataFrame 数据框 (以前 3 行为例):

python 复制代码 
print("[Message] Builded Spark DataFrame:")
SDF.show()

输出:

txt 复制代码 
+----------+-------+----+----+----+-----+---------+---------+-------------+----------+----+----+
|      Date|   Code|Open|High| Low|Close|Pre_Close|   Change|Turnover_Rate|    Volume| MA5|MA10|
+----------+-------+----+----+----+-----+---------+---------+-------------+----------+----+----+
|2015-12-31|'000422|7.93|7.95|7.76| 7.77|     7.93|-0.020177|     0.015498| 1.39152E7|7.86|7.85|
|2015-12-30|'000422|7.86|7.93|7.75| 7.93|     7.84|  0.01148|     0.018662| 1.67559E7|7.90|7.85|
|2015-12-29|'000422|7.72|7.85|7.69| 7.84|     7.71| 0.016861|     0.015886| 1.42638E7|7.90|7.81|
+----------+-------+----+----+----+-----+---------+---------+-------------+----------+----+----+

使用 Spark 读取 CSV 数据

调用 SparkSession 的 .read 方法读取 CSV 数据:

其中 .option 是读取文件时的选项, 左边是 "键(Key)", 右边是 "值(Value)", 例如 .option("header", "true") 与 {header = "true"} 类同。

python 复制代码 
SDF = spark.read.option("header", "true").option("encoding", "utf-8").csv("file:///D:\\HBYH_000422_20150806_20151231.csv")

输出 DataFrame 数据框:

python 复制代码 
print("[Message] Readed CSV File: D:\\HBYH_000422_20150806_20151231.csv")
SDF.show()

输出:

txt 复制代码 
[Message] Readed CSV File: D:\HBYH_000422_20150806_20151231.csv
+----------+-------+----+----+----+-----+---------+---------+-------------+--------+----+----+
|      Date|   Code|Open|High| Low|Close|Pre_Close|   Change|Turnover_Rate|  Volume| MA5|MA10|
+----------+-------+----+----+----+-----+---------+---------+-------------+--------+----+----+
|2015-12-31|'000422|7.93|7.95|7.76| 7.77|     7.93|-0.020177|     0.015498|13915200|7.86|7.85|
|2015-12-30|'000422|7.86|7.93|7.75| 7.93|     7.84| 0.011480|     0.018662|16755900|7.90|7.85|
|2015-12-29|'000422|7.72|7.85|7.69| 7.84|     7.71| 0.016861|     0.015886|14263800|7.90|7.81|
|2015-12-28|'000422|8.03|8.08|7.70| 7.71|     8.03|-0.039851|     0.030821|27672800|7.91|7.78|
|2015-12-25|'000422|8.03|8.05|7.93| 8.03|     7.99| 0.005006|     0.021132|18974000|7.93|7.78|
|2015-12-24|'000422|7.93|8.16|7.87| 7.99|     7.92| 0.008838|     0.026487|23781900|7.85|7.72|
|2015-12-23|'000422|7.97|8.11|7.88| 7.92|     7.89| 0.003802|     0.042360|38033600|7.80|7.69|
|2015-12-22|'000422|7.86|7.93|7.76| 7.89|     7.83| 0.007663|     0.026929|24178700|7.73|7.68|
|2015-12-21|'000422|7.59|7.89|7.56| 7.83|     7.63| 0.026212|     0.030777|27633600|7.66|7.67|
|2015-12-18|'000422|7.71|7.74|7.57| 7.63|     7.74|-0.014212|     0.024764|22234900|7.62|7.71|
|2015-12-17|'000422|7.58|7.75|7.57| 7.74|     7.55| 0.025166|     0.028054|25188400|7.59|7.77|
|2015-12-16|'000422|7.57|7.62|7.53| 7.55|     7.55| 0.000000|     0.020718|18601600|7.58|7.79|
|2015-12-15|'000422|7.63|7.66|7.52| 7.55|     7.62|-0.009186|     0.025902|23256600|7.64|7.78|
|2015-12-14|'000422|7.40|7.64|7.36| 7.62|     7.51| 0.014647|     0.021005|18860100|7.68|7.76|
|2015-12-11|'000422|7.65|7.70|7.41| 7.51|     7.67|-0.020860|     0.020477|18385900|7.80|7.73|
|2015-12-10|'000422|7.78|7.87|7.65| 7.67|     7.83|-0.020434|     0.019972|17931900|7.95|7.69|
|2015-12-09|'000422|7.76|8.00|7.75| 7.83|     7.77| 0.007722|     0.025137|22569700|8.00|7.68|
|2015-12-08|'000422|8.08|8.18|7.76| 7.77|     8.24|-0.057039|     0.036696|32948200|7.92|7.66|
|2015-12-07|'000422|8.12|8.39|7.94| 8.24|     8.23| 0.001215|     0.064590|57993100|7.84|7.64|
|2015-12-04|'000422|7.85|8.48|7.80| 8.23|     7.92| 0.039141|     0.100106|89881900|7.65|7.58|
+----------+-------+----+----+----+-----+---------+---------+-------------+--------+----+----+
only showing top 20 rows

转换 Spark 中 DateFrame 各列数据类型

通常情况下, 为了避免计算出现数据类型的错误, 都需要重新转换一下数据类型。

python 复制代码 
# 转换 Spark 中 DateFrame 数据类型。
SDF = SDF.withColumn("Date",          col("Date").cast(DateType()))
SDF = SDF.withColumn("Open",          col("Open").cast(DoubleType()))
SDF = SDF.withColumn("High",          col("High").cast(DoubleType()))
SDF = SDF.withColumn("Low",           col("Low").cast(DoubleType()))
SDF = SDF.withColumn("Close",         col("Close").cast(DoubleType()))
SDF = SDF.withColumn("Pre_Close",     col("Pre_Close").cast(DoubleType()))
SDF = SDF.withColumn("Change",        col("Change").cast(DoubleType()))
SDF = SDF.withColumn("Turnover_Rate", col("Turnover_Rate").cast(DoubleType()))
SDF = SDF.withColumn("Volume",        col("Volume").cast(IntegerType()))
SDF = SDF.withColumn("MA5",           col("MA5").cast(DoubleType()))
SDF = SDF.withColumn("MA10",          col("MA10").cast(DoubleType()))

# 输出 Spark 中 DataFrame 字段和数据类型。
print("[Message] Changed Spark DataFrame Data Type:")
SDF.printSchema()

输出:

txt 复制代码 
[Message] Changed Spark DataFrame Data Type:
root
 |-- Date: date (nullable = true)
 |-- Code: string (nullable = true)
 |-- Open: double (nullable = true)
 |-- High: double (nullable = true)
 |-- Low: double (nullable = true)
 |-- Close: double (nullable = true)
 |-- Pre_Close: double (nullable = true)
 |-- Change: double (nullable = true)
 |-- Turnover_Rate: double (nullable = true)
 |-- Volume: integer (nullable = true)
 |-- MA5: double (nullable = true)
 |-- MA10: double (nullable = true)

将 Spark 的 DateFrame 和 Spark RDD 互相转换并计算数据

编写 "向 spark.sql 的 Row 对象添加字段和字段值" 函数:

python 复制代码 
def MapFunc_SparkSQL_Row_Add_Field(SrcRow:pyspark.sql.types.Row, FldName:str, FldVal:object) -> pyspark.sql.types.Row: 

    """
    [Require] import pyspark
    
    [Example] >>> SrcRow = Row(Date=datetime.date(2023, 12, 1), Clerk='Bob', Incom=5432.10)
              >>> NewRow = MapFunc_SparkSQL_Row_Add_Field(SrcRow=SrcRow, FldName='Weekday', FldVal=SrcRow['Date'].weekday())
              >>> print(NewRow)
              Row(Date=datetime.date(2023, 12, 1), Clerk='Bob', Incom=5432.10, Weekday=4)
    """
    
    # Convert Obj "pyspark.sql.types.Row" to Dict. 
    # ----------------------------------------------
    Row_Dict = SrcRow.asDict()
    
    # Add a New Key in the Dictionary With the New Column Name and Value.
    # ----------------------------------------------
    Row_Dict[FldName] = FldVal
    
    # Convert Dict to Obj "pyspark.sql.types.Row". 
    # ----------------------------------------------
    NewRow = pyspark.sql.types.Row(**Row_Dict)
    
    # ==============================================
    return NewRow

编写 "判断股票涨跌" 函数:

python 复制代码 
def MapFunc_Stock_Judgement_Rise_or_Fall(ChgRate:float) -> int: 

    if (ChgRate >= 0.0): return 1
    if (ChgRate <  0.0): return 0
    
    # ==============================================
    # End of Function.

编写 "判断股票短期均线和长期均线关系" 函数:

python 复制代码 
def MapFunc_Stock_Judgement_Short_MA_and_Long_MA_Relationship(Short_MA:float, Long_MA:float) -> int: 

    if (Short_MA >= Long_MA): return  1
    if (Short_MA == Long_MA): return  0
    if (Short_MA <= Long_MA): return -1
    
    # ==============================================
    # End of Function.

编写 "返回星期几(中文)" 函数:

python 复制代码 
def DtmFunc_Weekday_Return_String_CN(SrcDtm:datetime.datetime) -> str:
    
    """
    [Require] import datetime
    
    [Explain] Python3 中 datetime.datetime 对象的 .weekday() 方法返回的是从 0 到 6 的数字 (0 代表周一, 6 代表周日)。
    """
    
    Weekday_Str_Chinese:list = ["周一", "周二", "周三", "周四", "周五", "周六", "周日"]
    
    # ==============================================
    return Weekday_Str_Chinese[SrcDtm.weekday()]

在 Spark 中将 DataFrame 转换为 Spark RDD 并调用自定义函数:

python 复制代码 
# 在 Spark 中将 DataFrame 转换为 RDD。
CalcRDD = SDF.rdd

# --------------------------------------------------
# 调用自定义函数: 提取星期索引。
CalcRDD = CalcRDD.map(lambda X: MapFunc_SparkSQL_Row_Add_Field(X, "Weekday(Idx)", X["Date"].weekday()))
# ..................................................
# 调用自定义函数: 返回星期几(中文)。
CalcRDD = CalcRDD.map(lambda X: MapFunc_SparkSQL_Row_Add_Field(X, "Weekday(CN)", DtmFunc_Weekday_Return_String_CN(X["Date"])))
# ..................................................
# 调用自定义函数: 判断股票涨跌。
CalcRDD = CalcRDD.map(lambda X: MapFunc_SparkSQL_Row_Add_Field(X, "Rise_Fall", MapFunc_Stock_Judgement_Rise_or_Fall(X["Change"])))
# ..................................................
# 判断股票短期均线和长期均线关系。
CalcRDD = CalcRDD.map(lambda X: MapFunc_SparkSQL_Row_Add_Field(X, "MA_Relationship", MapFunc_Stock_Judgement_Short_MA_and_Long_MA_Relationship(Short_MA=X["MA5"], Long_MA=X["MA10"])))

# 显示计算好的 RDD 前 5 行。
print("[Message] Calculated RDD Top 5 Rows:")
pprint.pprint(CalcRDD.take(5))

输出:

txt 复制代码 
[Message] Calculated RDD Top 5 Rows:
[Row(Date=datetime.date(2015, 12, 31), Code="'000422", Open=7.93, High=7.95, Low=7.76, Close=7.77, Pre_Close=7.93, Change=-0.020177, Turnover_Rate=0.015498, Volume=13915200, MA5=7.86, MA10=7.85, Weekday(Idx)=3, Weekday(CN)='周四', Rise_Fall=0, MA_Relationship=1),
 Row(Date=datetime.date(2015, 12, 30), Code="'000422", Open=7.86, High=7.93, Low=7.75, Close=7.93, Pre_Close=7.84, Change=0.01148, Turnover_Rate=0.018662, Volume=16755900, MA5=7.9, MA10=7.85, Weekday(Idx)=2, Weekday(CN)='周三', Rise_Fall=1, MA_Relationship=1),
 Row(Date=datetime.date(2015, 12, 29), Code="'000422", Open=7.72, High=7.85, Low=7.69, Close=7.84, Pre_Close=7.71, Change=0.016861, Turnover_Rate=0.015886, Volume=14263800, MA5=7.9, MA10=7.81, Weekday(Idx)=1, Weekday(CN)='周二', Rise_Fall=1, MA_Relationship=1),
 Row(Date=datetime.date(2015, 12, 28), Code="'000422", Open=8.03, High=8.08, Low=7.7, Close=7.71, Pre_Close=8.03, Change=-0.039851, Turnover_Rate=0.030821, Volume=27672800, MA5=7.91, MA10=7.78, Weekday(Idx)=0, Weekday(CN)='周一', Rise_Fall=0, MA_Relationship=1),
 Row(Date=datetime.date(2015, 12, 25), Code="'000422", Open=8.03, High=8.05, Low=7.93, Close=8.03, Pre_Close=7.99, Change=0.005006, Turnover_Rate=0.021132, Volume=18974000, MA5=7.93, MA10=7.78, Weekday(Idx)=4, Weekday(CN)='周五', Rise_Fall=1, MA_Relationship=1)]

计算完成后将 Spark RDD 转换回 Spark 的 DataFrame:

python 复制代码 
# 在 Spark 中将 RDD 转换为 DataFrame。
NewSDF = CalcRDD.toDF()

print("[Message] Convert RDD to DataFrame and Filter Out Key Columns for Display:")
NewSDF.select(["Date", "Code", "High", "Low", "Close", "Change", "MA5", "MA10", "Weekday(CN)", "Rise_Fall", "MA_Relationship"]).show()

输出:

txt 复制代码 
[Message] Convert RDD to DataFrame and Filter Out Key Columns:
+----------+-------+----+----+-----+---------+----+----+-----------+---------+---------------+
|      Date|   Code|High| Low|Close|   Change| MA5|MA10|Weekday(CN)|Rise_Fall|MA_Relationship|
+----------+-------+----+----+-----+---------+----+----+-----------+---------+---------------+
|2015-12-31|'000422|7.95|7.76| 7.77|-0.020177|7.86|7.85|       周四|        0|              1|
|2015-12-30|'000422|7.93|7.75| 7.93|  0.01148| 7.9|7.85|       周三|        1|              1|
|2015-12-29|'000422|7.85|7.69| 7.84| 0.016861| 7.9|7.81|       周二|        1|              1|
|2015-12-28|'000422|8.08| 7.7| 7.71|-0.039851|7.91|7.78|       周一|        0|              1|
|2015-12-25|'000422|8.05|7.93| 8.03| 0.005006|7.93|7.78|       周五|        1|              1|
|2015-12-24|'000422|8.16|7.87| 7.99| 0.008838|7.85|7.72|       周四|        1|              1|
|2015-12-23|'000422|8.11|7.88| 7.92| 0.003802| 7.8|7.69|       周三|        1|              1|
|2015-12-22|'000422|7.93|7.76| 7.89| 0.007663|7.73|7.68|       周二|        1|              1|
|2015-12-21|'000422|7.89|7.56| 7.83| 0.026212|7.66|7.67|       周一|        1|             -1|
|2015-12-18|'000422|7.74|7.57| 7.63|-0.014212|7.62|7.71|       周五|        0|             -1|
|2015-12-17|'000422|7.75|7.57| 7.74| 0.025166|7.59|7.77|       周四|        1|             -1|
|2015-12-16|'000422|7.62|7.53| 7.55|      0.0|7.58|7.79|       周三|        1|             -1|
|2015-12-15|'000422|7.66|7.52| 7.55|-0.009186|7.64|7.78|       周二|        0|             -1|
|2015-12-14|'000422|7.64|7.36| 7.62| 0.014647|7.68|7.76|       周一|        1|             -1|
|2015-12-11|'000422| 7.7|7.41| 7.51| -0.02086| 7.8|7.73|       周五|        0|              1|
|2015-12-10|'000422|7.87|7.65| 7.67|-0.020434|7.95|7.69|       周四|        0|              1|
|2015-12-09|'000422| 8.0|7.75| 7.83| 0.007722| 8.0|7.68|       周三|        1|              1|
|2015-12-08|'000422|8.18|7.76| 7.77|-0.057039|7.92|7.66|       周二|        0|              1|
|2015-12-07|'000422|8.39|7.94| 8.24| 0.001215|7.84|7.64|       周一|        1|              1|
|2015-12-04|'000422|8.48| 7.8| 8.23| 0.039141|7.65|7.58|       周五|        1|              1|
+----------+-------+----+----+-----+---------+----+----+-----------+---------+---------------+

字符串索引化 (StringIndexer) 演示 (Only Demo)

StringIndexer (字符串-索引变换) 是一个估计器, 是将字符串列编码为标签索引列。索引位于 [0, numLabels), 按标签频率排序, 频率最高的排 0, 依次类推, 因此最常见的标签获取索引是 0。

python 复制代码 
# 使用 StringIndexer 转换 Weekday(CN) 列。
MyStringIndexer = StringIndexer(inputCol="Weekday(CN)", outputCol="StrIdx")
# 拟合并转换数据。
IndexedSDF = MyStringIndexer.fit(NewSDF).transform(NewSDF)

# 筛选 Date, Weekday(Idx), Weekday(CN), StrIdx 四列, 输出 StringIndexer 效果。
print("[Message] The Effect of StringIndexer:")
IndexedSDF.select(["Date", "Weekday(Idx)", "Weekday(CN)", "StrIdx"]).show()

输出:

txt 复制代码 
[Message] The Effect of StringIndexer:
+----------+------------+-----------+------+
|      Date|Weekday(Idx)|Weekday(CN)|StrIdx|
+----------+------------+-----------+------+
|2015-12-31|           3|       周四|   3.0|
|2015-12-30|           2|       周三|   1.0|
|2015-12-29|           1|       周二|   2.0|
|2015-12-28|           0|       周一|   0.0|
|2015-12-25|           4|       周五|   4.0|
|2015-12-24|           3|       周四|   3.0|
|2015-12-23|           2|       周三|   1.0|
|2015-12-22|           1|       周二|   2.0|
|2015-12-21|           0|       周一|   0.0|
|2015-12-18|           4|       周五|   4.0|
|2015-12-17|           3|       周四|   3.0|
|2015-12-16|           2|       周三|   1.0|
|2015-12-15|           1|       周二|   2.0|
|2015-12-14|           0|       周一|   0.0|
|2015-12-11|           4|       周五|   4.0|
|2015-12-10|           3|       周四|   3.0|
|2015-12-09|           2|       周三|   1.0|
|2015-12-08|           1|       周二|   2.0|
|2015-12-07|           0|       周一|   0.0|
|2015-12-04|           4|       周五|   4.0|
+----------+------------+-----------+------+
only showing top 20 rows

提取 标签(Label)列 和 特征向量(Features)列

在创建特征向量(Features)列时, 将会用到 VectorAssembler 模块, VectorAssembler 将多个特征合并为一个特征向量。

提取 标签(Label) 列:

python 复制代码 
# 将 Rise_Fall 列复制为 Label 列。
NewSDF = NewSDF.withColumn("Label", col("Rise_Fall"))

创建 特征向量(Features) 列:

python 复制代码 
# VectorAssembler 将多个特征合并为一个特征向量。
FeaColsName:list = ["High", "Low", "Turnover_Rate", "Volume", "Weekday(Idx)", "MA_Relationship"]
MyAssembler = VectorAssembler(inputCols=FeaColsName, outputCol="Features")

# 拟合数据 (可选, 如果在模型训练时使用 Pipeline, 则无需在此步骤拟合数据, 当然也就无法在此步骤预览数据)。
AssembledSDF = MyAssembler.transform(NewSDF)

输出预览:

python 复制代码 
print("[Message] Assembled Label and Features for GBTClassifier:")
AssembledSDF.select(["Date", "Code", "High", "Low", "Close", "Change", "MA5", "MA10", "Weekday(CN)", "Rise_Fall", "MA_Relationship", "Label", "Features"]).show()

预览:

txt 复制代码 
[Message] Assembled for GBTClassifier:
+----------+-------+----+----+-----+---------+----+----+-----------+---------+---------------+-----+--------------------+
|      Date|   Code|High| Low|Close|   Change| MA5|MA10|Weekday(CN)|Rise_Fall|MA_Relationship|Label|            Features|
+----------+-------+----+----+-----+---------+----+----+-----------+---------+---------------+-----+--------------------+
|2015-12-31|'000422|7.95|7.76| 7.77|-0.020177|7.86|7.85|       周四|        0|              1|    0|[7.95,7.76,0.0154...|
|2015-12-30|'000422|7.93|7.75| 7.93|  0.01148| 7.9|7.85|       周三|        1|              1|    1|[7.93,7.75,0.0186...|
|2015-12-29|'000422|7.85|7.69| 7.84| 0.016861| 7.9|7.81|       周二|        1|              1|    1|[7.85,7.69,0.0158...|
|2015-12-28|'000422|8.08| 7.7| 7.71|-0.039851|7.91|7.78|       周一|        0|              1|    0|[8.08,7.7,0.03082...|
|2015-12-25|'000422|8.05|7.93| 8.03| 0.005006|7.93|7.78|       周五|        1|              1|    1|[8.05,7.93,0.0211...|
|2015-12-24|'000422|8.16|7.87| 7.99| 0.008838|7.85|7.72|       周四|        1|              1|    1|[8.16,7.87,0.0264...|
|2015-12-23|'000422|8.11|7.88| 7.92| 0.003802| 7.8|7.69|       周三|        1|              1|    1|[8.11,7.88,0.0423...|
|2015-12-22|'000422|7.93|7.76| 7.89| 0.007663|7.73|7.68|       周二|        1|              1|    1|[7.93,7.76,0.0269...|
|2015-12-21|'000422|7.89|7.56| 7.83| 0.026212|7.66|7.67|       周一|        1|             -1|    1|[7.89,7.56,0.0307...|
|2015-12-18|'000422|7.74|7.57| 7.63|-0.014212|7.62|7.71|       周五|        0|             -1|    0|[7.74,7.57,0.0247...|
|2015-12-17|'000422|7.75|7.57| 7.74| 0.025166|7.59|7.77|       周四|        1|             -1|    1|[7.75,7.57,0.0280...|
|2015-12-16|'000422|7.62|7.53| 7.55|      0.0|7.58|7.79|       周三|        1|             -1|    1|[7.62,7.53,0.0207...|
|2015-12-15|'000422|7.66|7.52| 7.55|-0.009186|7.64|7.78|       周二|        0|             -1|    0|[7.66,7.52,0.0259...|
|2015-12-14|'000422|7.64|7.36| 7.62| 0.014647|7.68|7.76|       周一|        1|             -1|    1|[7.64,7.36,0.0210...|
|2015-12-11|'000422| 7.7|7.41| 7.51| -0.02086| 7.8|7.73|       周五|        0|              1|    0|[7.7,7.41,0.02047...|
|2015-12-10|'000422|7.87|7.65| 7.67|-0.020434|7.95|7.69|       周四|        0|              1|    0|[7.87,7.65,0.0199...|
|2015-12-09|'000422| 8.0|7.75| 7.83| 0.007722| 8.0|7.68|       周三|        1|              1|    1|[8.0,7.75,0.02513...|
|2015-12-08|'000422|8.18|7.76| 7.77|-0.057039|7.92|7.66|       周二|        0|              1|    0|[8.18,7.76,0.0366...|
|2015-12-07|'000422|8.39|7.94| 8.24| 0.001215|7.84|7.64|       周一|        1|              1|    1|[8.39,7.94,0.0645...|
|2015-12-04|'000422|8.48| 7.8| 8.23| 0.039141|7.65|7.58|       周五|        1|              1|    1|[8.48,7.8,0.10010...|
+----------+-------+----+----+-----+---------+----+----+-----------+---------+---------------+-----+--------------------+
only showing top 20 rows

训练 梯度提升树分类器(GBTClassifier) 模型

将数据集划分为 "训练集" 和 "测试集":

python 复制代码 
(TrainingData, TestData) = AssembledSDF.randomSplit([0.8, 0.2], seed=42)

创建 梯度提升树分类器(GBTClassifier):

python 复制代码 
GBTC = GBTClassifier(labelCol="Label", featuresCol="Features", maxIter=10)

创建 Pipeline (可选):

python 复制代码 
# 创建 Pipeline, 将特征向量转换和梯度提升树模型组合在一起
# 注意: 如果要使用 Pipeline, 则在创建 特征向量(Features)列 的时候不需要拟合数据, 否则会报 "Output column Features already exists." 的错误。
MyPipeline = Pipeline(stages=[MyAssembler, GBTC])

训练 梯度提升树分类器(GBTClassifier) 模型:

如果在创建 特征向量(Features)列 的时候已经拟合数据:

python 复制代码 
# 训练模型 (普通模式)。
Model = GBTC.fit(TrainingData)

如果在创建 特征向量(Features)列 的时候没有拟合数据:

python 复制代码 
# 训练模型 (Pipeline 模式)。
Model = MyPipeline.fit(TrainingData)

使用 梯度提升树分类器(GBTClassifier) 模型预测数据

python 复制代码 
# 在测试集上进行预测。
Predictions = Model.transform(TestData)

# 删除不需要的列 (以免列数太多, 结果显示拥挤, 不好观察)。
Predictions = Predictions.drop("Open")
Predictions = Predictions.drop("High")
Predictions = Predictions.drop("Low")
Predictions = Predictions.drop("Close")
Predictions = Predictions.drop("Pre_Close")
Predictions = Predictions.drop("Turnover_Rate")
Predictions = Predictions.drop("Volume")
Predictions = Predictions.drop("Weekday(Idx)")
Predictions = Predictions.drop("Weekday(CN)")

print("[Message] Prediction Results on The Test Data Set for GBTClassifier:")
Predictions.show()

输出:

txt 复制代码 
[Message] Prediction Results on The Test Data Set for GBTClassifier:
+----------+-------+---------+----+----+---------+---------------+-----+--------------------+--------------------+--------------------+----------+
|      Date|   Code|   Change| MA5|MA10|Rise_Fall|MA_Relationship|Label|            Features|       rawPrediction|         probability|prediction|
+----------+-------+---------+----+----+---------+---------------+-----+--------------------+--------------------+--------------------+----------+
|2015-08-10|'000422| 0.034105| 8.2|7.92|        1|              1|    1|[8.58,8.18,0.0412...|[-1.2336852238184...|[0.07817752585370...|       1.0|
|2015-08-14|'000422| 0.009479|8.43|8.24|        1|              1|    1|[8.65,8.43,0.0411...|[0.87088297782943...|[0.85091123576658...|       0.0|
|2015-08-18|'000422|-0.095455|8.39|8.32|        0|              1|    0|[8.86,7.92,0.0561...|[-1.2043889179555...|[0.08250578702314...|       1.0|
|2015-08-25|'000422|-0.099424|7.52|7.96|        0|             -1|    0|[6.77,6.25,0.0294...|[-1.0533113737445...|[0.10845478968034...|       1.0|
|2015-09-02|'000422|-0.053412|6.73|6.91|        0|             -1|    0|[6.88,6.3,0.02228...|[-1.1111838911434...|[0.09775976029530...|       1.0|
|2015-09-10|'000422|-0.031161|6.76|6.74|        0|              1|    0|[7.01,6.76,0.0174...|[-0.4697765236086...|[0.28099063380343...|       1.0|
|2015-09-18|'000422|      0.0|6.39|6.62|        1|             -1|    1|[6.58,6.3,0.01662...|[-0.5683790237136...|[0.24291608452764...|       1.0|
|2015-09-28|'000422| 0.009464|6.48|6.47|        1|              1|    1|[6.42,6.25,0.0088...|[-1.0881365069305...|[0.10190150697248...|       1.0|
|2015-10-19|'000422|-0.007062|6.94|6.72|        0|              1|    0|[7.13,6.92,0.0312...|[-1.0878476739737...|[0.10195438572358...|       1.0|
|2015-10-20|'000422| 0.008535|6.98|6.81|        1|              1|    1|[7.09,6.94,0.0244...|[-1.0333458639919...|[0.11237660949811...|       1.0|
|2015-10-21|'000422|-0.062059|6.96|6.85|        0|              1|    0|[7.11,6.61,0.0393...|[-1.2022113775327...|[0.08283606045630...|       1.0|
|2015-10-23|'000422| 0.054412|6.95|6.93|        1|              1|    1|[7.22,6.81,0.0471...|[-1.2563395161683...|[0.07497409890575...|       1.0|
|2015-10-27|'000422| 0.033426|7.04|7.01|        1|              1|    1|[7.48,7.08,0.0576...|[-1.1731833621455...|[0.08735499814440...|       1.0|
|2015-11-02|'000422|-0.027548|7.23| 7.1|        0|              1|    0|[7.26,7.05,0.0168...|[-0.8503605266136...|[0.15437111484969...|       1.0|
|2015-11-11|'000422| 0.005284|7.54|7.37|        1|              1|    1|[7.64,7.52,0.0261...|[-0.3579712978325...|[0.32828707596815...|       1.0|
|2015-11-20|'000422| 0.002635|7.52|7.53|        1|             -1|    1|[7.71,7.53,0.0282...|[-0.0263480243045...|[0.48682903554857...|       1.0|
|2015-12-02|'000422| 0.009511|7.37|7.49|        1|             -1|    1|[7.48,7.2,0.01596...|[0.95178733740680...|[0.87029557357980...|       0.0|
+----------+-------+---------+----+----+---------+---------------+-----+--------------------+--------------------+--------------------+----------+

使用 BinaryClassificationEvaluator 评估模型性能

python 复制代码 
# 使用 BinaryClassificationEvaluator 评估模型性能。
MyEvaluator = BinaryClassificationEvaluator(labelCol="Label", metricName="areaUnderROC")
auc = MyEvaluator.evaluate(Predictions)

print("Area Under ROC (AUC):", auc)

输出:

txt 复制代码 
Area Under ROC (AUC): 0.37142857142857144

完整代码

python 复制代码 
#!/usr/bin/python3
# Create By GF 2024-01-07

# 在这个例子中, 我们使用 VectorAssembler 将多个特征列合并为一个特征向量, 并使用 GBTClassifier 构建梯度提升树模型。
# 最后, 我们使用 BinaryClassificationEvaluator 评估模型性能, 通常使用 ROC 曲线下面积 (AUC) 作为评估指标。
# 请根据你的实际数据和问题调整特征列, 标签列以及其他参数。在实际应用中, 你可能需要进行更多的特征工程, 调参和模型评估。

import datetime
import pprint
# --------------------------------------------------
import pyspark
# --------------------------------------------------
from pyspark.sql import Row, SparkSession
from pyspark.sql.functions import col
from pyspark.sql.types import DateType, IntegerType, DoubleType
# --------------------------------------------------
from pyspark.ml.feature import StringIndexer, VectorAssembler
from pyspark.ml.classification import GBTClassifier
from pyspark.ml.evaluation import BinaryClassificationEvaluator
from pyspark.ml import Pipeline

# 编写 "向 spark.sql 的 Row 对象添加字段和字段值" 函数。
def MapFunc_SparkSQL_Row_Add_Field(SrcRow:pyspark.sql.types.Row, FldName:str, FldVal:object) -> pyspark.sql.types.Row: 

    """
    [Require] import pyspark
    
    [Example] >>> SrcRow = Row(Date=datetime.date(2023, 12, 1), Clerk='Bob', Incom=5432.10)
              >>> NewRow = MapFunc_SparkSQL_Row_Add_Field(SrcRow=SrcRow, FldName='Weekday', FldVal=SrcRow['Date'].weekday())
              >>> print(NewRow)
              Row(Date=datetime.date(2023, 12, 1), Clerk='Bob', Incom=5432.10, Weekday=4)
    """
    
    # Convert Obj "pyspark.sql.types.Row" to Dict. 
    # ----------------------------------------------
    Row_Dict = SrcRow.asDict()
    
    # Add a New Key in the Dictionary With the New Column Name and Value.
    # ----------------------------------------------
    Row_Dict[FldName] = FldVal
    
    # Convert Dict to Obj "pyspark.sql.types.Row". 
    # ----------------------------------------------
    NewRow = pyspark.sql.types.Row(**Row_Dict)
    
    # ==============================================
    return NewRow

# 编写 "判断股票涨跌" 函数。
def MapFunc_Stock_Judgement_Rise_or_Fall(ChgRate:float) -> int: 

    if (ChgRate >= 0.0): return 1
    if (ChgRate <  0.0): return 0
    
    # ==============================================
    # End of Function.

# 编写 "判断股票短期均线和长期均线关系" 函数。
def MapFunc_Stock_Judgement_Short_MA_and_Long_MA_Relationship(Short_MA:float, Long_MA:float) -> int: 

    if (Short_MA >= Long_MA): return  1
    if (Short_MA == Long_MA): return  0
    if (Short_MA <= Long_MA): return -1
    
    # ==============================================
    # End of Function.

# 编写 "返回星期几(中文)" 函数。
def DtmFunc_Weekday_Return_String_CN(SrcDtm:datetime.datetime) -> str:
    
    """
    [Require] import datetime
    
    [Explain] Python3 中 datetime.datetime 对象的 .weekday() 方法返回的是从 0 到 6 的数字 (0 代表周一, 6 代表周日)。
    """
    
    Weekday_Str_Chinese:list = ["周一", "周二", "周三", "周四", "周五", "周六", "周日"]
    
    # ==============================================
    return Weekday_Str_Chinese[SrcDtm.weekday()]

if __name__ == "__main__":

    # Spark 2.0 以上版本的 spark-shell 在启动时会自动创建一个名为 spark 的 SparkSession 对象。
    # 当需要手工创建时, SparkSession 可以由其伴生对象的 builder 方法创建出来。
    spark = SparkSession.builder.master("local[*]").appName("spark").getOrCreate()
    
    # 调用 SparkSession 的 .read 方法读取 CSV 数据:
    # 其中 .option 是读取文件时的选项, 左边是 "键(Key)", 右边是 "值(Value)", 例如 .option("header", "true") 与 {header = "true"} 类同。
    SDF = spark.read.option("header", "true").option("encoding", "utf-8").csv("file:///D:\\HBYH_000422_20150806_20151231.csv")
    
    print("[Message] Readed CSV File: D:\\HBYH_000422_20150806_20151231.csv")
    SDF.show()
    
    # 转换 Spark 中 DateFrame 数据类型。
    SDF = SDF.withColumn("Date",          col("Date").cast(DateType()))
    SDF = SDF.withColumn("Open",          col("Open").cast(DoubleType()))
    SDF = SDF.withColumn("High",          col("High").cast(DoubleType()))
    SDF = SDF.withColumn("Low",           col("Low").cast(DoubleType()))
    SDF = SDF.withColumn("Close",         col("Close").cast(DoubleType()))
    SDF = SDF.withColumn("Pre_Close",     col("Pre_Close").cast(DoubleType()))
    SDF = SDF.withColumn("Change",        col("Change").cast(DoubleType()))
    SDF = SDF.withColumn("Turnover_Rate", col("Turnover_Rate").cast(DoubleType()))
    SDF = SDF.withColumn("Volume",        col("Volume").cast(IntegerType()))
    SDF = SDF.withColumn("MA5",           col("MA5").cast(DoubleType()))
    SDF = SDF.withColumn("MA10",          col("MA10").cast(DoubleType()))
    
    # 输出 Spark 中 DataFrame 字段和数据类型。
    print("[Message] Changed Spark DataFrame Data Type:")
    SDF.printSchema()
    
    # 在 Spark 中将 DataFrame 转换为 RDD。
    CalcRDD = SDF.rdd
    
    # --------------------------------------------------
    # 调用自定义函数: 提取星期索引。
    CalcRDD = CalcRDD.map(lambda X: MapFunc_SparkSQL_Row_Add_Field(X, "Weekday(Idx)", X["Date"].weekday()))
    # ..................................................
    # 调用自定义函数: 返回星期几(中文)。
    CalcRDD = CalcRDD.map(lambda X: MapFunc_SparkSQL_Row_Add_Field(X, "Weekday(CN)", DtmFunc_Weekday_Return_String_CN(X["Date"])))
    # ..................................................
    # 调用自定义函数: 判断股票涨跌。
    CalcRDD = CalcRDD.map(lambda X: MapFunc_SparkSQL_Row_Add_Field(X, "Rise_Fall", MapFunc_Stock_Judgement_Rise_or_Fall(X["Change"])))
    # ..................................................
    # 判断股票短期均线和长期均线关系。
    CalcRDD = CalcRDD.map(lambda X: MapFunc_SparkSQL_Row_Add_Field(X, "MA_Relationship", MapFunc_Stock_Judgement_Short_MA_and_Long_MA_Relationship(Short_MA=X["MA5"], Long_MA=X["MA10"])))
    
    # 显示计算好的 RDD 前 5 行。
    print("[Message] Calculated RDD Top 5 Rows:")
    pprint.pprint(CalcRDD.take(5))
    
    # 在 Spark 中将 RDD 转换为 DataFrame。
    NewSDF = CalcRDD.toDF()
    
    print("[Message] Convert RDD to DataFrame and Filter Out Key Columns for Display:")
    NewSDF.select(["Date", "Code", "High", "Low", "Close", "Change", "MA5", "MA10", "Weekday(CN)", "Rise_Fall", "MA_Relationship"]).show()
    
    # 提取 标签(Label) 列: 将 Rise_Fall 列复制为 Label 列。
    NewSDF = NewSDF.withColumn("Label", col("Rise_Fall"))
    
    # 创建 特征向量(Features) 列: VectorAssembler 将多个特征合并为一个特征向量。
    FeaColsName:list = ["High", "Low", "Turnover_Rate", "Volume", "Weekday(Idx)", "MA_Relationship"]
    MyAssembler = VectorAssembler(inputCols=FeaColsName, outputCol="Features")

    # 创建 特征向量(Features) 列: 拟合数据 (可选, 如果在模型训练时使用 Pipeline, 则无需在此步骤拟合数据, 当然也就无法在此步骤预览数据)。
    AssembledSDF = MyAssembler.transform(NewSDF)
    
    print("[Message] Assembled Label and Features for GBTClassifier:")
    AssembledSDF.select(["Date", "Code", "High", "Low", "Close", "Change", "MA5", "MA10", "Weekday(CN)", "Rise_Fall", "MA_Relationship", "Label", "Features"]).show()
    
    # 将数据集划分为 "训练集" 和 "测试集"。
    (TrainingData, TestData) = AssembledSDF.randomSplit([0.8, 0.2], seed=42)
    
    # 创建 梯度提升树分类器(GBTClassifier)。
    GBTC = GBTClassifier(labelCol="Label", featuresCol="Features", maxIter=10)
    
    # 创建 Pipeline (可选): 将特征向量转换和梯度提升树模型组合在一起
    # 注意: 如果要使用 Pipeline, 则在创建 特征向量(Features)列 的时候不需要拟合数据, 否则会报 "Output column Features already exists." 的错误。
    #MyPipeline = Pipeline(stages=[MyAssembler, GBTC])
    
    # 训练模型 (普通模式)。
    Model = GBTC.fit(TrainingData)
    
    # 训练模型 (Pipeline 模式)。
    #Model = MyPipeline.fit(TrainingData)
    
    # 在测试集上进行预测。
    Predictions = Model.transform(TestData)
    
    # 删除不需要的列 (以免列数太多, 结果显示拥挤, 不好观察)。
    Predictions = Predictions.drop("Open")
    Predictions = Predictions.drop("High")
    Predictions = Predictions.drop("Low")
    Predictions = Predictions.drop("Close")
    Predictions = Predictions.drop("Pre_Close")
    Predictions = Predictions.drop("Turnover_Rate")
    Predictions = Predictions.drop("Volume")
    Predictions = Predictions.drop("Weekday(Idx)")
    Predictions = Predictions.drop("Weekday(CN)")
    
    print("[Message] Prediction Results on The Test Data Set for GBTClassifier:")
    Predictions.show()
    
    # 使用 BinaryClassificationEvaluator 评估模型性能。
    MyEvaluator = BinaryClassificationEvaluator(labelCol="Label", metricName="areaUnderROC")
    auc = MyEvaluator.evaluate(Predictions)

    print("Area Under ROC (AUC):", auc)

其它

在这个例子中, 我们使用 VectorAssembler 将多个特征列合并为一个特征向量, 并使用 GBTClassifier 构建梯度提升树模型。

最后, 我们使用 BinaryClassificationEvaluator 评估模型性能, 通常使用 ROC 曲线下面积 (AUC) 作为评估指标。

请根据你的实际数据和问题调整特征列, 标签列以及其他参数。在实际应用中, 你可能需要进行更多的特征工程, 调参和模型评估。

总结

以上就是关于 金融数据 PySpark-3.0.3梯度提升树(GBTClassifier)实例 的全部内容。

更多内容可以访问我的代码仓库:

https://gitee.com/goufeng928/public

https://github.com/goufeng928/public

相关推荐
正在走向自律2 分钟前
微众银行金融场景 Agent：创新实践与深度剖析（12/30）
金融·ai agent·ai智能体·微众银行
刚学HTML7 分钟前
leetcode 05 回文字符串
算法·leetcode
AC使者26 分钟前
#B1630. 数字走向4
算法
ROBOT玲玉27 分钟前
Milvus 中，FieldSchema 的 dim 参数和索引参数中的 “nlist“ 的区别
python·机器学习·numpy
冠位观测者30 分钟前
【Leetcode 每日一题】2545. 根据第 K 场考试的分数排序
数据结构·算法·leetcode
GocNeverGiveUp37 分钟前
机器学习2-NumPy
人工智能·机器学习·numpy
古希腊掌管学习的神1 小时前
[搜广推]王树森推荐系统笔记——曝光过滤 & Bloom Filter
算法·推荐算法
qystca1 小时前
洛谷 P1706 全排列问题 C语言
算法
浊酒南街1 小时前
决策树（理论知识1）
算法·决策树·机器学习
就爱学编程2 小时前
重生之我在异世界学编程之C语言小项目：通讯录
c语言·开发语言·数据结构·算法
热门推荐
01【一文读懂】NTN（非地面网络）技术介绍02优化手机性能，解决卡顿问题：关闭这3个微信开关，释放内存空间03玄机平台应急响应—webshell查杀04Cube-Studio：开源大模型全链路一站式中台05基于大语言模型（LLM）的合成数据生成、策展和评估的综述06(欧拉)openEuler系统添加网卡文件配置流程、(欧拉)openEuler系统手动配置ipv6地址流程、(欧拉)openEuler系统网络管理说明07组基轨迹建模 GBTM的介绍与实现（Stata 或 R）08HCIA-datacom数通题库和录播视频资料09Docker 夺命连环 15 问10RAG 实践- Ollama+RagFlow 部署本地知识库