시계열 스터디(7주차) : 주식 종가 예측하기 (내부 Contest)

 

---

 

1. Contest 내용

 

• 주요 내용 
  
  : 주어진 종목과 기간에 맞추어 시계열 모델을 제작해 주식 종가 가격 예측하기
  
  
•  규칙
  
  1) 종목 종류 : 삼성전자 (005930), 카카오 (035720), 하이브(352820) / 더네이쳐홀딩스(298540), 코엔텍(029960)
  
  2) 기간 : 훈련(22-11-01 ~ 23-04-21) / 예측 (23-04-24 ~ 23-04-28)

 

참고 : 데이콘 주식 종료가격 예측 대회

월간 데이콘 주식 종료 가격 예측 Pre Competiton (with.데이커) - DACON

 

2.  진행 내용

 

2-1. 데이터 불러오기 및 전처리

 

!pip install finance-datareader

import FinanceDataReader as fdr
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from tqdm import tqdm_notebook
import tqdm
from sklearn.linear_model import LinearRegression
from sklearn.ensemble import RandomForestRegressor
from sklearn.linear_model import LinearRegression, HuberRegressor
from sklearn.model_selection import GridSearchCV
import xgboost as xgb

# 삼성전자(005930), 카카오 (035720), 하이브(352820), 더네이쳐홀딩스(298540), 코엔텍(029960)
samsung = fdr.DataReader('005930', '2022-11-01', '2023-04-21')
kakao = fdr.DataReader('035720', '2022-11-01', '2023-04-21')
hive = fdr.DataReader('352820', '2022-11-01', '2023-04-21')
nature = fdr.DataReader('298540', '2022-11-01', '2023-04-21')
cnt = fdr.DataReader('029960', '2022-11-01', '2023-04-21')

start_date = '20221101'
end_date = '20230421'
stock_code = ['005930', '035720', '352820', '298540', '029960']

start_weekday = pd.to_datetime(start_date).weekday()
max_weeknum = pd.to_datetime(end_date).strftime('%V')
Business_days = pd.DataFrame(pd.date_range(start_date,end_date,freq='B'), columns = ['Date'])

print(f'WEEKDAY of "start_date" : {start_weekday}')
print(f'NUM of WEEKS to "end_date" : {max_weeknum}')
print(f'HOW MANY "Business_days" : {Business_days.shape}', )
display(Business_days.head())

 

NMAE 점수 계산 함수

def get_prediction(x_close, y_close, x_close_public):
    prediction_close = 0.0
    for i, model in enumerate(models):
        model.fit(x_close, y_close)
        prediction_close += model.predict(np.expand_dims(x_close_public, 0))*models_rate[i]
    return prediction_close
    
# 예측 결과 저장
pred_date = [20230424,20230425,20230426,20230427,20230428]
pred_table = pd.DataFrame({'Day' : pred_date})

 

EDA 결과 trend는 딱히 없어 추가 전처리를 하지 않고 진행 (모든 종목에서 적용)

 

2-2.  선형 회귀, Random Forest, XgBoost

 

기본 Logic : 월 ~ 금 데이터로 다음주 월 ~ 금 을 차례로 예측하여 주 단위 예측 진행

def NMAE(y_pred, y_train):
    return np.mean(abs(y_train-y_pred)/y_train)*100

model1 = LinearRegression(n_jobs=-1)
model2 = RandomForestRegressor(criterion="mae")
model3 = xgb.XGBRegressor(n_jobs=-1, eval_metric=NMAE)
models = [model1, model2, model3]

model1_rate = 0.5
model2_rate = 0.4
model3_rate = 0.1
models_rate = [model1_rate, model2_rate, model3_rate]

for code in stock_code:
    data = fdr.DataReader(code, start = start_date, end = end_date)[['Close']].reset_index()
    data = pd.merge(Business_days, data, how = 'outer')
    data['weekday'] = data.Date.apply(lambda x : x.weekday())
    data['weeknum'] = data.Date.apply(lambda x : x.strftime('%V'))
    data.Close = data.Close.ffill()
    data.Close = data.Close.bfill()

    data_close = pd.pivot_table(data = data, values = 'Close', columns = 'weekday', index = 'weeknum')

    data_close_flatten = np.ravel(data_close.to_numpy())

    predictions = []

    
    # day 1

    data_close_day1 = data_close_flatten.reshape((-1,1))
    data_close_day1 = data_close_day1[np.logical_not(np.isnan(data_close_day1))]

    x_close = data_close_day1[:-6]  
    y_close = data_close_day1[1:-5] 
    x_close_public = data_close_day1[-6]

    x_close = x_close.reshape((-1,1))
    y_close = y_close.reshape((-1,1))
    x_close_public = x_close_public.reshape((-1,1))

    prediction_close = get_prediction(x_close, y_close, x_close_public)
    predictions.append(prediction_close)
    
    
    # day 2
    

    data_close_day2 = np.insert(data_close_flatten, 0, data_close_flatten[0])
    data_close_day2 = data_close_day2[:-1]
    data_close_day2 = data_close_day2.reshape((-1,2))
    data_close_day2 = data_close_day2[np.logical_not(np.isnan(data_close_day2))]

    x_close = data_close_day2[:-3]
    y_close = data_close_day2[1:-2][:, 1]  
    x_close_public = data_close_day2[-3]

    prediction_close = get_prediction(x_close, y_close, x_close_public)
    predictions.append(prediction_close)
    

    # day 3
    
    data_close_day3 = data_close_flatten[1:]
    data_close_day3 = np.append(data_close_day3, 0)
    data_close_day3 = data_close_day3.reshape((-1, 3))
    data_close_day3 = data_close_day3[np.logical_not(np.isnan(data_close_day3))]

    x_close = data_close_day3[:-3]
    y_close = data_close_day3[1:-2][:, 2]
    x_close_public = data_close_day3[-3]
        
    prediction_close = get_prediction(x_close, y_close, x_close_public)
    predictions.append(prediction_close)

    
    # day 4
    
    data_close_day4 = np.insert(data_close_flatten, 0, data_close_flatten[0])
    data_close_day4 = data_close_day4[:-1]
    data_close_day4 = data_close_day4.reshape((-1, 4))
    data_close_day4 = data_close_day4[np.logical_not(np.isnan(data_close_day4))]

    x_close = data_close_day4[:-2]
    y_close = data_close_day4[1:-1][:, 3]
    x_close_public = data_close_day4[-2]

    prediction_close = get_prediction(x_close, y_close, x_close_public)
    predictions.append(prediction_close)
    
    # day 5
    
    data_close_day5 = data_close_flatten.reshape((-1, 5))
    data_close_day5 = data_close_day5[np.logical_not(np.isnan(data_close_day5))]

    x_close = data_close_day5[:-2]
    y_close = data_close_day5[1:-1][:, 4]
    x_close_public = data_close_day5[-2]

    prediction_close = get_prediction(x_close, y_close, x_close_public)
    predictions.append(prediction_close)
    
        
    pred_table.loc[:5, stock_code] = predictions
pred_table.isna().sum().sum()

 

<성능>

• Linear Regression : 1.4356
• RandomForest : 1.133
• XgBoost : 1.113

 

2-3. ARIMA

EDA로 ACF, PACF를 그린 결과 종목에 상관없이 다음과 같이 유사한 절단점을 보임.

 

from statsmodels.tsa.arima_model import ARIMA
import statsmodels.api as sm


data.index = data['Date']
data.drop('Date', axis=1, inplace=True)
data.columns = stock_code
data.fillna(method='bfill', inplace=True)

# Public
train = data[:-15]
for col in train.columns:
  model = ARIMA(train[col].values, order = (0,1,1)) 
  model_fit = model.fit(trend = 'c', full_output = True, disp = True)
  forecast_data = model_fit.forecast(steps=5)
  pred_arima_y = forecast_data[0].tolist()   
  pred_table[col][:5] = pred_arima_y

# Private
train = data
for col in train.columns:
  model = ARIMA(train[col].values, order = (0,1,1))
  model_fit = model.fit(trend = 'c', full_output = True, disp = True)
  forecast_data = model_fit.forecast(steps=5)
  pred_arima_y = forecast_data[0].tolist()   
  pred_table[col][5:] = pred_arima_y

 

<성능>

• ARIMA : 1.104

 

2-4. Prophet

위와 동일하게 period = 5 (다음주 영업일 예측)로 설정했지만 Prophet은 period = 20(다음달 영업일 예측) 성능이 더 좋게 나왔음.

# Period = 5

for i in stock_code:
    globals()['m_{}'.format(i)] = Prophet()
    locals()['m_{}'.format(i)].fit(locals()['df_{}'.format(i)])
    
    globals()['future_prices_{}'.format(i)] = locals()['m_{}'.format(i)].make_future_dataframe(periods=5)
    globals()['forecast_{}'.format(i)] = locals()['m_{}'.format(i)].predict(locals()['future_prices_{}'.format(i)])
    globals()['fore_cast_{}'.format(i)] = locals()['forecast_{}'.format(i)][['ds', 'yhat']]
    locals()['fore_cast_{}'.format(i)].rename(columns={'yhat': i}, inplace=True)
    locals()['fore_cast_{}'.format(i)].rename(columns={'ds': 'Day'}, inplace=True)
    sample_submission = pd.merge(sample_submission, locals()['fore_cast_{}'.format(i)], how='inner', on='Day')

 

# Period = 21

for i in stock_code:
    globals()['m_{}'.format(i)] = Prophet()
    locals()['m_{}'.format(i)].fit(locals()['df_{}'.format(i)])
    
    globals()['future_prices_{}'.format(i)] = locals()['m_{}'.format(i)].make_future_dataframe(periods=21)
    globals()['forecast_{}'.format(i)] = locals()['m_{}'.format(i)].predict(locals()['future_prices_{}'.format(i)])
    globals()['fore_cast_{}'.format(i)] = locals()['forecast_{}'.format(i)][['ds', 'yhat']]
    locals()['fore_cast_{}'.format(i)].rename(columns={'yhat': i}, inplace=True)
    locals()['fore_cast_{}'.format(i)].rename(columns={'ds': 'Day'}, inplace=True)
    sample_submission = pd.merge(sample_submission, locals()['fore_cast_{}'.format(i)], how='inner', on='Day')

 

<성능>

• Prophet(period = 5) : 1.222
• Prophet(period = 21) : 1.109

 

Prophet은 단기간 예측보다는 장기간 period 설정이 더 좋은 효과를 내는 것 같음.

 

https://facebook.github.io/prophet/docs/non-daily_data.html

Non-Daily Data

 

 

---

<Ref>

• https://dacon.io/competitions/official/235800/codeshare/3248?page=2&dtype=recent
• https://dacon.io/competitions/official/235800/codeshare/3249?page=2&dtype=recent