-
Notifications
You must be signed in to change notification settings - Fork 0
/
preprocess.py
389 lines (311 loc) · 14.8 KB
/
preprocess.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
import os
import random
from enum import Enum
import pandas as pd
import numpy as np
from category_encoders import BinaryEncoder
from sklearn.preprocessing import OrdinalEncoder, RobustScaler
from sklearn.base import TransformerMixin
from sklearn.feature_selection import SelectKBest, f_classif
from sklearn.impute import SimpleImputer
from sklearn.model_selection import train_test_split
from tqdm import tqdm
from pickle import dump, load
from pandas.tseries.offsets import MonthEnd
import yfinance as yf
import matplotlib.pyplot as plt
from tensorflow.keras.preprocessing.sequence import pad_sequences
"""
Idea:
Takes N most recent QRs as input and outputs buy, sell or hold for next quarter
- if number of most recent QRs < N, then data is padded with zeros from the start
- if number of most recent QRs > N, then data is truncated from the start (oldest)
- assumes future performance is not always independent from the past (rejects random walk hypothesis?)
- this is a multivariate times series classification
# - takes relative change as input but also considers relative total assets (not the change thereof)
# - assumes relative size of a company may matter too
- if there exists data for a company over a period of N quarters than N-2 different training data can be generated
using a sliding window method (potential data leak?)
"""
class StockClass(Enum):
HOLD = 0
BUY = 1
SELL = 2
class OutlierNullifier(TransformerMixin):
def __init__(self, **kwargs):
"""
Create a transformer to remove outliers.
Returns:
object: to be used as a transformer method as part of Pipeline()
"""
self.quantiles = {}
def fit(self, X, y=None, **fit_params):
if isinstance(X, np.ndarray):
for i in range(X.shape[1]):
self.quantiles[i] = np.quantile(X[i], [0.25, 0.75])
else:
for column in X.columns:
self.quantiles[column] = X[column].quantile(0.25), X[column].quantile(0.75)
return self
def transform(self, X, y=None):
if isinstance(X, np.ndarray):
for i in range(X.shape[1]):
q1, q3 = self.quantiles[i]
iqr = q3 - q1
X[i] = np.where((X[i] < q1 - 1.5 * iqr) | (X[i] > q3 + 1.5 * iqr), np.nan, X[i])
else:
for column in X.columns:
q1, q3 = self.quantiles[column]
iqr = q3 - q1
X[column] = np.where((X[column] < q1 - 1.5 * iqr) | (X[column] > q3 + 1.5 * iqr), np.nan, X[column])
return X
def fit_transform(self, X, y=None, **fit_params):
return self.fit(X, y, **fit_params).transform(X, y)
def set_index_to_date(df):
"""
Returns a sorted datetime index
"""
df['Quarter end'] = pd.to_datetime(df['Quarter end'])
df = df.set_index("Quarter end")
return df.sort_index(ascending=True)
def class_creation(df, threshold=0.03):
"""
Creates classes of:
- hold(0)
- buy(1)
- sell(2)
Threshold can be changed to fit whatever price percentage change is desired
"""
if df['Price high'] >= threshold and df['Price low'] >= threshold:
# Buys
return StockClass.BUY.value
elif df['Price high'] <= -threshold and df['Price low'] <= -threshold:
# Sells
return StockClass.SELL.value
else:
# Holds
return StockClass.HOLD.value
def save_as(obj, path):
os.makedirs(os.path.dirname(path), exist_ok=True)
with open(path, 'wb') as file:
dump(obj, file)
def save_as_x_y(data, X_path, y_path):
y = {k: df.pop('Label') for k, df in data.items()}
X = data
save_as(X, X_path)
save_as(y, y_path)
def clean(df):
# Convert dates to datetime
df['Quarter end'] = pd.to_datetime(df['Quarter end'], errors='coerce').dt.date
# Replace missing flags with np.nan
df['Stock'] = np.where((df['Stock'] == 'None') | (df['Stock'] == ''), np.nan, df['Stock'])
# Drop invalid rows
df = df.dropna(subset=['Stock', 'Quarter end'], how='any')
# Set and sort multi-index
df = df.set_index(['Stock', 'Quarter end'])
df = df.sort_index(level=df.index.names)
# Remove duplicates
duplicates = df.index.duplicated(keep='first')
df = df[~duplicates]
# Convert numeric data to numeric data
df = df.apply(pd.to_numeric, errors='coerce')
# # todo: Insert missing timestamps
# for stock in tqdm(df.index.get_level_values('Stock').unique()):
# timestamps = df.xs(stock).index
# idx = pd.period_range(min(timestamps), max(timestamps))
# df.loc[stock, :] = df.loc[stock, :].reindex(idx, fill_value=0)
return df
class StockFundamentalDataImputer(TransformerMixin):
def __init__(self, columns=None, method='linear', limit_direction='both'):
self.columns = columns
self.limit_direction = limit_direction
self.method = method
self.imputer = SimpleImputer()
def fit(self, X, y=None):
if self.columns is None:
self.columns = X.columns
self.imputer = self.imputer.fit(X[self.columns])
return self
def transform(self, X):
for stock in tqdm(X['Stock'].unique()):
subset = X['Stock'] == stock
# numeric_columns = X.drop(columns=['Stock', 'Quarter end']).columns
numeric_columns = X.drop(columns=['Stock']).columns
# Imputation via interpolation
X.loc[subset, numeric_columns] = X.loc[subset, numeric_columns].interpolate(
method=self.method, limit_direction=self.limit_direction, axis=0) # spline may be better?
X[self.columns] = pd.DataFrame(self.imputer.transform(X[self.columns]), columns=self.columns, index=X.index)
return X
def engineer_features(df, add_stock_info=False, classification=True):
for stock in tqdm(df['Stock'].unique()):
subset = df['Stock'] == stock
# numeric_columns = df.drop(columns=['Stock', 'Quarter end']).columns
numeric_columns = df.drop(columns=['Stock']).columns
# Replace values with percent difference or change
df.loc[subset, numeric_columns] = df.loc[subset, numeric_columns].pct_change(periods=1)
# Replace infinite values and nan with 0
df.loc[subset, numeric_columns] = df.loc[subset, numeric_columns].replace([np.inf, -np.inf], 0)
df.loc[subset, numeric_columns] = df.loc[subset, numeric_columns].fillna(0)
# Create the class 'Label' determining if a quarterly reports improvement is a buy, hold, or sell.
# shifted by -1 to know if the prices will increase/decrease in the next quarter
if classification:
df.loc[subset, 'Label'] = df[subset].apply(class_creation, axis=1).shift(-1)
else:
df.loc[subset, 'Label'] = df.loc[subset, 'Price'].shift(-1)
# Exclude the first and last rows (cannot label)
df = df.drop(index=[df[subset].index[0], df[subset].index[-1]])
# Add additional company info as features (very slow)
if add_stock_info:
info = yf.Ticker(stock).info
company_info = ['industry', 'sector', 'country', 'market']
for col in company_info:
value = info.get(col)
value = ['N/A'] if value is None else value
df.loc[subset, col] = value
# Drop the price related columns to prevent data leakage
df = df.drop(['Price', 'Price high', 'Price low'], axis=1)
return df
# Deviation Augmentation
def augment(df, sigma=0.05, size=20):
scalars = np.random.normal(1, sigma, size)
result = pd.DataFrame()
for scalar in scalars:
new_df = train_data.copy()
new_df['Stock'] = new_df['Stock'] + str(scalar)
# numeric_columns = [col for col in df.columns if col not in ['Stock', 'Quarter end']]
numeric_columns = [col for col in df.columns if col != 'Stock']
new_df[numeric_columns] = new_df[numeric_columns] * scalar
result = pd.concat([result, new_df])
return result
class PadTruncateTransformer(TransformerMixin):
def __init__(self, maxlen=None, padding='pre', truncating='pre', dtype='float'):
self.length = None
self.padding = padding
self.truncating = truncating
self.dtype = dtype
def fit(self, X, y=None, quantile=0.9):
self.length = int(np.quantile(X['Stock'].value_counts(), quantile))
return self
def transform(self, X):
# X = pd.DataFrame(pad_sequences(X.values, padding=self.padding, truncating=self.truncating, dtype=self.dtype), columns=X.columns)
X = self._pad_data(X, self.length, self.padding, self.truncating)
return X
def _pad_data(self, df, length, padding, truncating):
"""
Transforms all dataframes within the data to a fixed length via padding or truncating. If padding, pad with 0s and
Hold for the label.
:param df: dataframe
:param length: target row count for the dataframes
:param padding: {'pre', 'post'} if 'pre' then pad from the start; if 'post' pad from the end
:param truncating: {'pre', 'post'} if 'pre' then truncate from the start; if 'post' truncate from the end
:return: a dictionary of dataframes indexed by date
"""
# assert isinstance(df, dict)
assert padding in ['pre', 'post'] and truncating in ['pre', 'post']
segments = []
for stock in tqdm(df['Stock'].unique()):
df_subset = df[df['Stock'] == stock]
padding_length = length - len(df_subset.index)
if padding_length > 0:
# pad
df_to_pad = pd.DataFrame({col: [0.0] * padding_length for col in df_subset.columns})
if padding == 'post':
segments.append(df_subset)
segments.append(df_to_pad)
else:
segments.append(df_to_pad)
segments.append(df_subset)
elif padding_length < 0:
# truncate
df_subset = df_subset[:padding_length] if truncating == 'post' else df_subset[-padding_length:]
segments.append(df_subset)
else:
segments.append(df_subset)
result = pd.concat(segments)
result = result.reset_index(drop=True)
return result
# desired_length = 101
if __name__ == '__main__':
# Parameters
use_augmentation = True # Augment training data via variance scaling, may cause data leak
# Load companies quarterly reports
try:
df = pd.read_csv('datasets/historical_qrs.csv')
df = clean(df)
except Exception:
df = pd.read_csv('datasets/historical_qrs.csv')
df = clean(df)
df.to_csv('datasets/clean_historical_qrs.csv', index=True)
# Split training set, test set and validation set
train_stocks, test_stocks = train_test_split(df['Stock'].unique(), test_size=0.2, random_state=42)
train_data, test_data = df[df['Stock'].isin(train_stocks)], df[df['Stock'].isin(test_stocks)]
train_stocks, val_stocks = train_test_split(train_data['Stock'].unique(), test_size=0.25, random_state=42)
train_data, val_data = train_data[train_data['Stock'].isin(train_stocks)], \
train_data[train_data['Stock'].isin(val_stocks)]
# Imputation
# imputer = StockFundamentalDataImputer(train_data.drop(columns=['Stock', 'Quarter end']).columns)
imputer = StockFundamentalDataImputer(train_data.drop(columns=['Stock']).columns)
train_data = imputer.fit_transform(train_data)
test_data = imputer.transform(test_data)
val_data = imputer.transform(val_data)
# Feature engineering
train_data = engineer_features(train_data, add_stock_info=False, classification=False)
test_data = engineer_features(test_data, add_stock_info=False, classification=False)
val_data = engineer_features(val_data, add_stock_info=False, classification=False)
# Data augmentation
# if use_augmentation:
# train_data = augment(train_data)
# Remove outliers
# outlier_remover = OutlierNullifier()
# features = train_data.drop(columns=['Label', 'Stock']).columns
# train_data[features] = outlier_remover.fit_transform(train_data[features])
# test_data[features] = outlier_remover.transform(test_data[features])
# val_data[features] = outlier_remover.transform(val_data[features])
# imputer = StockFundamentalDataImputer(train_data.drop(columns=['Stock']).columns)
# train_data = imputer.fit_transform(train_data)
# test_data = imputer.transform(test_data)
# val_data = imputer.transform(val_data)
# Encode
encoder = OrdinalEncoder(handle_unknown='use_encoded_value', unknown_value=-1)
train_data['Stock'] = encoder.fit_transform(train_data[['Stock']])
test_data['Stock'] = encoder.transform(test_data[['Stock']])
val_data['Stock'] = encoder.transform(val_data[['Stock']])
# Scale
scaler = RobustScaler()
features = train_data.drop(columns=['Label']).columns
train_data[features] = scaler.fit_transform(train_data[features])
test_data[features] = scaler.transform(test_data[features])
val_data[features] = scaler.transform(val_data[features])
# Feature selection
# features = train_data.drop(columns=['Label', 'Stock', 'Quarter end']).columns
# features = train_data.drop(columns=['Label', 'Stock']).columns
# selector = SelectKBest(f_classif, 30)
# val_data[features] = selector.fit_transform(val_data[features], val_data['Label'])
# train_data[features] = selector.transform(train_data[features])
# test_data[features] = selector.transform(test_data[features])
# Pad data
padder = PadTruncateTransformer(padding='pre', truncating='pre', dtype='float')
train_data = padder.fit_transform(train_data)
test_data = padder.transform(test_data)
val_data = padder.transform(val_data)
# Save data
train_data.to_csv('datasets/train_data.csv', index=False)
test_data.to_csv('datasets/test_data.csv', index=False)
val_data.to_csv('datasets/val_data.csv', index=False)
# Extract X, y
y_train, X_train = train_data.pop('Label'), train_data
y_test, X_test = test_data.pop('Label'), test_data
y_val, X_val = val_data.pop('Label'), val_data
# Reshape as ndarrays (n_stocks, n_timestamps, n_features)
X_train = X_train.values.reshape(-1, padder.length, len(X_train.columns))
y_train = y_train.values.reshape(-1, padder.length, 1)
X_test = X_test.values.reshape(-1, padder.length, len(X_test.columns))
y_test = y_test.values.reshape(-1, padder.length, 1)
X_val = X_val.values.reshape(-1, padder.length, len(X_val.columns))
y_val = y_val.values.reshape(-1, padder.length, 1)
save_as(X_train, 'datasets/X_train.pkl')
save_as(y_train, 'datasets/y_train.pkl')
save_as(X_test, 'datasets/X_test.pkl')
save_as(y_test, 'datasets/y_test.pkl')
save_as(X_val, 'datasets/X_val.pkl')
save_as(y_val, 'datasets/y_val.pkl')