Create model_90m.py

model_90m.py

@@ -0,0 +1,399 @@
+import streamlit as st
+import pandas as pd
+import pandas_datareader as pdr
+import numpy as np
+import yfinance as yf
+import json
+import requests
+from bs4 import BeautifulSoup
+from typing import List
+import xgboost as xgb
+from tqdm import tqdm
+from sklearn import linear_model
+import joblib
+import os
+from sklearn.metrics import roc_auc_score, precision_score, recall_score
+import datetime
+from pandas.tseries.offsets import BDay
+from datasets import load_dataset
+
+def walk_forward_validation(df, target_column, num_training_rows, num_periods):
+    
+    # Create an XGBRegressor model
+    # model = xgb.XGBRegressor(n_estimators=100, objective='reg:squarederror', random_state = 42)
+    model = linear_model.LinearRegression()
+
+    overall_results = []
+    # Iterate over the rows in the DataFrame, one step at a time
+    for i in tqdm(range(num_training_rows, df.shape[0] - num_periods + 1),desc='LR Model'):
+        # Split the data into training and test sets
+        X_train = df.drop(target_column, axis=1).iloc[:i]
+        y_train = df[target_column].iloc[:i]
+        X_test = df.drop(target_column, axis=1).iloc[i:i+num_periods]
+        y_test = df[target_column].iloc[i:i+num_periods]
+        
+        # Fit the model to the training data
+        model.fit(X_train, y_train)
+        
+        # Make a prediction on the test data
+        predictions = model.predict(X_test)
+        
+        # Create a DataFrame to store the true and predicted values
+        result_df = pd.DataFrame({'True': y_test, 'Predicted': predictions}, index=y_test.index)
+        
+        overall_results.append(result_df)
+
+    df_results = pd.concat(overall_results)
+    # model.save_model('model_lr.bin')
+    # Return the true and predicted values, and fitted model
+    return df_results, model
+
+def walk_forward_validation_seq(df, target_column_clf, target_column_regr, num_training_rows, num_periods):
+
+    # Create run the regression model to get its target
+    res, model1 = walk_forward_validation(df.drop(columns=[target_column_clf]).dropna(), target_column_regr, num_training_rows, num_periods)
+    # joblib.dump(model1, 'model1.bin')
+
+    # Merge the result df back on the df for feeding into the classifier
+    for_merge = res[['Predicted']]
+    for_merge.columns = ['RegrModelOut']
+    for_merge['RegrModelOut'] = for_merge['RegrModelOut'] > 0
+    df = df.merge(for_merge, left_index=True, right_index=True)
+    df = df.drop(columns=[target_column_regr])
+    df = df[[
+        'CurrentGap','RegrModelOut',
+        'CurrentHigh30toClose',
+        'CurrentLow30toClose',
+        'CurrentClose30toClose',
+        'CurrentRange30',
+        'GapFill30',target_column_clf
+        ]]
+    
+    df[target_column_clf] = df[target_column_clf].astype(bool)
+    df['RegrModelOut'] = df['RegrModelOut'].astype(bool)
+
+    # Create an XGBRegressor model
+    model2 = xgb.XGBClassifier(n_estimators=10, random_state = 42)
+    # model = linear_model.LogisticRegression(max_iter=1500)
+    
+    overall_results = []
+    # Iterate over the rows in the DataFrame, one step at a time
+    for i in tqdm(range(num_training_rows, df.shape[0] - num_periods + 1),'CLF Model'):
+        # Split the data into training and test sets
+        X_train = df.drop(target_column_clf, axis=1).iloc[:i]
+        y_train = df[target_column_clf].iloc[:i]
+        X_test = df.drop(target_column_clf, axis=1).iloc[i:i+num_periods]
+        y_test = df[target_column_clf].iloc[i:i+num_periods]
+        
+        # Fit the model to the training data
+        model2.fit(X_train, y_train)
+        
+        # Make a prediction on the test data
+        predictions = model2.predict_proba(X_test)[:,-1]
+        
+        # Create a DataFrame to store the true and predicted values
+        result_df = pd.DataFrame({'True': y_test, 'Predicted': predictions}, index=y_test.index)
+        
+        overall_results.append(result_df)
+
+    df_results = pd.concat(overall_results)
+    # model1.save_model('model_ensemble.bin')
+    # joblib.dump(model2, 'model2.bin')
+    # Return the true and predicted values, and fitted model
+    return df_results, model1, model2
+
+def seq_predict_proba(df, trained_reg_model, trained_clf_model):
+    regr_pred = trained_reg_model.predict(df)
+    regr_pred = regr_pred > 0
+    new_df = df.copy()
+    new_df['RegrModelOut'] = regr_pred
+    clf_pred_proba = trained_clf_model.predict_proba(new_df[['CurrentGap','RegrModelOut',
+    'CurrentHigh30toClose',
+    'CurrentLow30toClose',
+    'CurrentClose30toClose',
+    'CurrentRange30',
+    'GapFill30']])[:,-1]
+    return clf_pred_proba
+
+def get_data():
+    # f = open('settings.json')
+    # j = json.load(f)
+    # API_KEY_FRED = j["API_KEY_FRED"]
+
+    API_KEY_FRED = os.getenv('API_KEY_FRED')
+    
+    def parse_release_dates(release_id: str) -> List[str]:
+        release_dates_url = f'https://api.stlouisfed.org/fred/release/dates?release_id={release_id}&realtime_start=2015-01-01&include_release_dates_with_no_data=true&api_key={API_KEY_FRED}'
+        r = requests.get(release_dates_url)
+        text = r.text
+        soup = BeautifulSoup(text, 'xml')
+        dates = []
+        for release_date_tag in soup.find_all('release_date', {'release_id': release_id}):
+            dates.append(release_date_tag.text)
+        return dates
+
+    def parse_release_dates_obs(series_id: str) -> List[str]:
+        obs_url = f'https://api.stlouisfed.org/fred/series/observations?series_id={series_id}&realtime_start=2015-01-01&include_release_dates_with_no_data=true&api_key={API_KEY_FRED}'
+        r = requests.get(obs_url)
+        text = r.text
+        soup = BeautifulSoup(text, 'xml')
+        observations  = []
+        for observation_tag in soup.find_all('observation'):
+            date = observation_tag.get('date')
+            value = observation_tag.get('value')
+            observations.append((date, value))
+        return observations
+
+    econ_dfs = {}
+
+    econ_tickers = [
+        'WALCL',
+        'NFCI',
+        'WRESBAL'
+    ]
+
+    for et in tqdm(econ_tickers, desc='getting econ tickers'):
+        # p = parse_release_dates_obs(et)
+        # df = pd.DataFrame(columns = ['ds',et], data = p)
+        df = pdr.get_data_fred(et)
+        df.index = df.index.rename('ds')
+        # df.index = pd.to_datetime(df.index.rename('ds')).dt.tz_localize(None)
+        # df['ds'] = pd.to_datetime(df['ds']).dt.tz_localize(None)
+        econ_dfs[et] = df
+
+    # walcl = pd.DataFrame(columns = ['ds','WALCL'], data = p)
+    # walcl['ds'] = pd.to_datetime(walcl['ds']).dt.tz_localize(None)
+
+    # nfci = pd.DataFrame(columns = ['ds','NFCI'], data = p2)
+    # nfci['ds'] = pd.to_datetime(nfci['ds']).dt.tz_localize(None)
+
+    release_ids = [
+        "10", # "Consumer Price Index"
+        "46", # "Producer Price Index"
+        "50", # "Employment Situation"
+        "53", # "Gross Domestic Product"
+        "103", # "Discount Rate Meeting Minutes"
+        "180", # "Unemployment Insurance Weekly Claims Report"
+        "194", # "ADP National Employment Report"
+        "323" # "Trimmed Mean PCE Inflation Rate"
+    ]
+
+    release_names = [
+        "CPI",
+        "PPI",
+        "NFP",
+        "GDP",
+        "FOMC",
+        "UNEMP",
+        "ADP",
+        "PCE"
+    ]
+
+    releases = {}
+
+    for rid, n in tqdm(zip(release_ids, release_names), total = len(release_ids), desc='Getting release dates'):
+        releases[rid] = {}
+        releases[rid]['dates'] = parse_release_dates(rid)
+        releases[rid]['name'] = n 
+
+    # Create a DF that has all dates with the name of the col as 1
+    # Once merged on the main dataframe, days with econ events will be 1 or None. Fill NA with 0
+    # This column serves as the true/false indicator of whether there was economic data released that day.
+    for rid in tqdm(release_ids, desc='Making indicators'):
+        releases[rid]['df'] = pd.DataFrame(
+            index=releases[rid]['dates'],
+            data={
+            releases[rid]['name']: 1
+            })
+        releases[rid]['df'].index = pd.DatetimeIndex(releases[rid]['df'].index)
+        # releases[rid]['df']['ds'] = pd.to_datetime(releases[rid]['df']['ds']).dt.tz_localize(None)
+        # releases[rid]['df'] = releases[rid]['df'].set_index('ds')
+
+    vix = yf.Ticker('^VIX')
+    spx = yf.Ticker('^GSPC')
+
+        
+    # Pull in data
+    data = load_dataset("boomsss/SPX_full_30min", split='train')
+
+    rows = [d['text'] for d in data]
+    rows = [x.split(',') for x in rows]
+
+    fr = pd.DataFrame(columns=[
+        'Datetime','Open','High','Low','Close'
+    ], data = rows)
+
+    fr['Datetime'] = pd.to_datetime(fr['Datetime'])
+    fr['Datetime'] = fr['Datetime'].dt.tz_localize('America/New_York')
+    fr = fr.set_index('Datetime')
+    fr['Open'] = pd.to_numeric(fr['Open'])
+    fr['High'] = pd.to_numeric(fr['High'])
+    fr['Low'] = pd.to_numeric(fr['Low'])
+    fr['Close'] = pd.to_numeric(fr['Close'])
+
+    # Get incremental date
+    last_date = fr.index.date[-1]
+    last_date = last_date + datetime.timedelta(days=1)
+    # Get incremental data
+    spx1 = yf.Ticker('^GSPC')
+    yfp = spx1.history(start=last_date, interval='90m')
+    # Concat current and incremental
+    df_30m = pd.concat([fr, yfp])
+    # Get the first 30 minute bar
+    df_30m = df_30m.reset_index()
+    df_30m['Datetime'] = df_30m['Datetime'].dt.date
+    df_30m = df_30m.groupby('Datetime').head(3)
+    df_30m = df_30m.set_index('Datetime',drop=True)
+    # Rename the columns
+    df_30m = df_30m[['Open','High','Low','Close']]
+
+    opens_1h = df_30m.groupby('Datetime')['Open'].head(1)
+    closes_1h = df_30m.groupby('Datetime')['Close'].tail(1)
+    highs_1h = df_30m.groupby('Datetime')['High'].max()
+    lows_1h = df_30m.groupby('Datetime')['Low'].min()
+    
+    df_1h = pd.DataFrame(index=df_30m.index.unique())
+    df_1h['Open'] = opens_1h
+    df_1h['Close'] = closes_1h
+    df_1h['High'] = highs_1h
+    df_1h['Low'] = lows_1h
+
+    df_1h.columns = ['Open30','High30','Low30','Close30']
+
+    prices_vix = vix.history(start='2018-07-01', interval='1d')
+    prices_spx = spx.history(start='2018-07-01', interval='1d')
+    prices_spx['index'] = [str(x).split()[0] for x in prices_spx.index]
+    prices_spx['index'] = pd.to_datetime(prices_spx['index']).dt.date
+    prices_spx.index = prices_spx['index']
+    prices_spx = prices_spx.drop(columns='index')
+    prices_spx.index = pd.DatetimeIndex(prices_spx.index)
+
+
+    prices_vix['index'] = [str(x).split()[0] for x in prices_vix.index]
+    prices_vix['index'] = pd.to_datetime(prices_vix['index']).dt.date
+    prices_vix.index = prices_vix['index']
+    prices_vix = prices_vix.drop(columns='index')
+    prices_vix.index = pd.DatetimeIndex(prices_vix.index)
+
+
+    data = prices_spx.merge(df_1h, left_index=True, right_index=True)
+    data = data.merge(prices_vix[['Open','High','Low','Close']], left_index=True, right_index=True, suffixes=['','_VIX'])
+
+    # Features
+    data['PrevClose'] = data['Close'].shift(1)
+    data['Perf5Day'] = data['Close'] > data['Close'].shift(5)
+    data['Perf5Day_n1'] = data['Perf5Day'].shift(1)
+    data['Perf5Day_n1'] = data['Perf5Day_n1'].astype(bool)
+    data['GreenDay'] = (data['Close'] > data['PrevClose']) * 1
+    data['RedDay'] = (data['Close'] <= data['PrevClose']) * 1
+
+    data['VIX5Day'] = data['Close_VIX'] > data['Close_VIX'].shift(5)
+    data['VIX5Day_n1'] = data['VIX5Day'].astype(bool)
+
+    data['Range'] = data[['Open','High']].max(axis=1) - data[['Low','Open']].min(axis=1) # Current day range in points
+    data['RangePct'] = data['Range'] / data['Close']
+    data['VIXLevel'] = pd.qcut(data['Close_VIX'], 4)
+    data['OHLC4_VIX'] = data[['Open_VIX','High_VIX','Low_VIX','Close_VIX']].mean(axis=1)
+    data['OHLC4'] = data[['Open','High','Low','Close']].mean(axis=1)
+    data['OHLC4_Trend'] = data['OHLC4'] > data['OHLC4'].shift(1)
+    data['OHLC4_Trend_n1'] = data['OHLC4_Trend'].shift(1)
+    data['OHLC4_Trend_n1'] = data['OHLC4_Trend_n1'].astype(float)
+    data['OHLC4_Trend_n2'] = data['OHLC4_Trend'].shift(1)
+    data['OHLC4_Trend_n2'] = data['OHLC4_Trend_n2'].astype(float)
+    data['RangePct_n1'] = data['RangePct'].shift(1)
+    data['RangePct_n2'] = data['RangePct'].shift(2)
+    data['OHLC4_VIX_n1'] = data['OHLC4_VIX'].shift(1)
+    data['OHLC4_VIX_n2'] = data['OHLC4_VIX'].shift(2)
+    data['CurrentGap'] = (data['Open'] - data['PrevClose']) / data['PrevClose']
+    data['CurrentGap'] = data['CurrentGap'].shift(-1)
+    data['DayOfWeek'] = pd.to_datetime(data.index)
+    data['DayOfWeek'] = data['DayOfWeek'].dt.day
+
+    # Intraday features
+    data['CurrentHigh30'] = data['High30'].shift(-1)
+    data['CurrentLow30'] = data['Low30'].shift(-1)
+    data['CurrentClose30'] = data['Close30'].shift(-1)
+
+    # Open to High
+    data['CurrentHigh30toClose'] = (data['CurrentHigh30'] / data['Close']) - 1
+    data['CurrentLow30toClose'] = (data['CurrentLow30'] / data['Close']) - 1
+    data['CurrentClose30toClose'] = (data['CurrentClose30'] / data['Close']) - 1
+    data['CurrentRange30'] = (data['CurrentHigh30'] - data['CurrentLow30']) / data['Close']
+    data['GapFill30'] = [low <= prev_close if gap > 0 else high >= prev_close for high, low, prev_close, gap in zip(data['CurrentHigh30'], data['CurrentLow30'], data['Close'], data['CurrentGap'])]
+
+    # Target -- the next day's low
+    data['Target'] = (data['OHLC4'] / data['PrevClose']) - 1
+    data['Target'] = data['Target'].shift(-1)
+    # data['Target'] = data['RangePct'].shift(-1)
+
+    # Target for clf -- whether tomorrow will close above or below today's close
+    data['Target_clf'] = data['Close'] > data['PrevClose']
+    data['Target_clf'] = data['Target_clf'].shift(-1)
+    data['DayOfWeek'] = pd.to_datetime(data.index)
+    data['Quarter'] = data['DayOfWeek'].dt.quarter
+    data['DayOfWeek'] = data['DayOfWeek'].dt.weekday
+
+    for rid in tqdm(release_ids, desc='Merging econ data'):
+        # Get the name of the release
+        n = releases[rid]['name']
+        # Merge the corresponding DF of the release
+        data = data.merge(releases[rid]['df'], how = 'left', left_index=True, right_index=True)
+        # Create a column that shifts the value in the merged column up by 1
+        data[f'{n}_shift'] = data[n].shift(-1)
+        # Fill the rest with zeroes
+        data[n] = data[n].fillna(0)
+        data[f'{n}_shift'] = data[f'{n}_shift'].fillna(0)
+        
+    data['BigNewsDay'] = data[[x for x in data.columns if '_shift' in x]].max(axis=1)
+
+    def cumul_sum(col):
+        nums = []
+        s = 0
+        for x in col:
+            if x == 1:
+                s += 1
+            elif x == 0:
+                s = 0
+            nums.append(s)
+        return nums
+
+    consec_green = cumul_sum(data['GreenDay'].values)
+    consec_red = cumul_sum(data['RedDay'].values)
+
+    data['DaysGreen'] = consec_green
+    data['DaysRed'] = consec_red
+
+    final_row = data.index[-2]
+
+    exp_row = data.index[-1]
+
+    df_final = data.loc[:final_row,
+    [
+        'BigNewsDay',
+        'Quarter',
+        'Perf5Day',
+        'Perf5Day_n1',
+        'DaysGreen',
+        'DaysRed',
+        'CurrentHigh30toClose',
+        'CurrentLow30toClose',
+        'CurrentClose30toClose',
+        'CurrentRange30',
+        'GapFill30',
+        # 'OHLC4_Trend',
+        # 'OHLC4_Trend_n1',
+        # 'OHLC4_Trend_n2',
+        # 'VIX5Day',
+        # 'VIX5Day_n1',
+        'CurrentGap',
+        'RangePct',
+        'RangePct_n1',
+        'RangePct_n2',
+        'OHLC4_VIX',
+        'OHLC4_VIX_n1',
+        'OHLC4_VIX_n2',
+        'Target',
+        'Target_clf'
+        ]]
+    df_final = df_final.dropna(subset=['Target','Target_clf','Perf5Day_n1'])
+    return data, df_final, final_row