85 KiB
85 KiB
Лабораторная работа 3¶
Датасет - Цены на кофе https://www.kaggle.com/datasets/mayankanand2701/starbucks-stock-price-dataset
Бизнес-цели:
- Прогнозирование цены закрытия акции для поддержки принятия решений по инвестициям.
- Оценка волатильности акций Starbucks для долгосрочных стратегий инвестирования.
Цели технического проекта:
- Создание модели машинного обучения для прогнозирования цены закрытия акций на основе исторических данных (дат, цен открытия, максимальных и минимальных цен, объёма торгов).
- Разработка системы, которая вычисляет и анализирует волатильность на основе исторической ценовой информации и объёмов торгов.
In [13]:
import pandas as pd
from sklearn.model_selection import train_test_split
from imblearn.under_sampling import RandomUnderSampler
df = pd.read_csv("data/Coffe.csv")
print(df)
df['Date'] = pd.to_datetime(df['Date'])
df['Date_numeric'] = (df['Date'] - pd.Timestamp('1970-01-01')).dt.days
print(df['Date_numeric'])
In [4]:
def split_stratified_into_train_val_test(
df_input,
stratify_colname="y",
frac_train=0.6,
frac_val=0.15,
frac_test=0.25,
random_state=None,
):
if frac_train + frac_val + frac_test != 1.0:
raise ValueError(
"fractions %f, %f, %f do not add up to 1.0"
% (frac_train, frac_val, frac_test)
)
if stratify_colname not in df_input.columns:
raise ValueError("%s is not a column in the dataframe" % (stratify_colname))
X = df_input
y = df_input[
[stratify_colname]
]
df_train, df_temp, y_train, y_temp = train_test_split(
X, y, stratify=y, test_size=(1.0 - frac_train), random_state=random_state
)
relative_frac_test = frac_test / (frac_val + frac_test)
df_val, df_test, y_val, y_test = train_test_split(
df_temp,
y_temp,
stratify=y_temp,
test_size=relative_frac_test,
random_state=random_state,
)
assert len(df_input) == len(df_train) + len(df_val) + len(df_test)
return df_train, df_val, df_test
bins = [df['Close'].min(), df['Close'].quantile(0.33), df['Close'].quantile(0.66), df['Close'].max()]
labels = ['Low', 'Medium', 'High']
df['Close_binned'] = pd.cut(df['Close'], bins=bins, labels=labels)
df = df.dropna()
# Now stratify using the binned values
df_train, df_val, df_test = split_stratified_into_train_val_test(
df, stratify_colname="Close_binned", frac_train=0.60, frac_val=0.20, frac_test=0.20
)
print(df_train.columns)
print("Обучающая выборка: ", df_train.shape)
print(df_train.Close.value_counts())
print("Контрольная выборка: ", df_val.shape)
print(df_val.Close.value_counts())
print("Тестовая выборка: ", df_test.shape)
print(df_test.Close.value_counts())
print("Обучающая выборка: ", df_train.shape)
print(df_train['Close_binned'].value_counts())
print("Контрольная выборка: ", df_val.shape)
print(df_val['Close_binned'].value_counts())
print("Тестовая выборка: ", df_test.shape)
print(df_test['Close_binned'].value_counts())
In [5]:
rus = RandomUnderSampler(random_state=42)
X_resampled, y_resampled = rus.fit_resample(df_train, df_train["Close_binned"])
df_train_rus = pd.DataFrame(X_resampled)
print("Обучающая выборка после undersampling: ", df_train_rus.shape)
print(df_train_rus.Close.value_counts())
In [6]:
df_train = pd.get_dummies(df_train, columns=['Close_binned'])
df_train['Volume_binned'] = pd.qcut(df_train['Volume'], q=4, labels=False)
df_train['Price_change'] = df_train['Close'] - df_train['Open']
print(df_train)
In [7]:
from sklearn.preprocessing import StandardScaler
scaler = StandardScaler()
df_train[['Open', 'Close', 'High', 'Low', 'Volume']] = scaler.fit_transform(
df_train[['Open', 'Close', 'High', 'Low', 'Volume']])
df_train['Volatility'] = df_train['High'] - df_train['Low']
print(df_train)
In [8]:
import featuretools as ft
es = ft.EntitySet(id="stocks")
es = es.add_dataframe(
dataframe_name="stock_data",
dataframe=df_train,
index="Date")
feature_matrix, feature_defs = ft.dfs(
entityset=es,
target_dataframe_name="stock_data")
feature_defs
Out[8]:
In [9]:
# Оценка предсказательной способности
from sklearn.linear_model import LinearRegression
from sklearn.metrics import mean_absolute_error, mean_squared_error
df_train_regression = df_train.copy()
X_train = df_train_regression.drop(['Close', 'Date'], axis=1)
y_train = df_train_regression['Close']
X_test = df_test.drop(['Close', 'Date'], axis=1)
y_test = df_test['Close']
X_train_encoded = pd.get_dummies(X_train, drop_first=True)
X_test_encoded = pd.get_dummies(X_test, drop_first=True)
X_test_encoded = X_test_encoded.reindex(columns=X_train_encoded.columns, fill_value=0)
print(X_train_encoded.dtypes)
In [10]:
model = LinearRegression()
model.fit(X_train_encoded, y_train)
predictions = model.predict(X_test_encoded)
mae = mean_absolute_error(y_test, predictions)
mse = mean_squared_error(y_test, predictions)
print("Средняя абсолютная ошибка:", mae)
print("Среднеквадратичная ошибка:", mse)
In [11]:
# Оценка скорости вычисления
import time
start_time = time.time()
model.fit(X_train_encoded, y_train)
training_time = time.time() - start_time
start_time = time.time()
predictions = model.predict(X_test_encoded)
prediction_time = time.time() - start_time
print(f'время, затраченное на обучение модели: {training_time}. Время, затраченное на предсказание: {prediction_time}')
In [12]:
# Оценка корреляции
import seaborn as sns
import matplotlib.pyplot as plt
corr_matrix = df_train_regression.corr()
sns.heatmap(corr_matrix, annot=False)
plt.show()
