virgo-modules 0.0.72__py3-none-any.whl → 0.8.4__py3-none-any.whl

virgo_modules/src/aws_utils.py

@@ -6,14 +6,36 @@ import pandas as pd
 
 
 def upload_file_to_aws(bucket,key,input_path, aws_credentials):
-    
+    '''
+    upload file from a folder to an s3 folder
+
+            Parameters:
+                    bucket (str): bucket name
+                    key (str): key pattern or folder in s3 e.g. path/to/upload/
+                    input_path (str): input path of the file to upload e.g. path/to/upload.txt
+                    aws_credentials (dict): aws credentials dictionary
+                    
+            Returns:
+                    None
+    '''
     session = boto3.Session(aws_access_key_id=aws_credentials['AWS_ACCESS_KEY_ID'],aws_secret_access_key=aws_credentials['AWS_SECRET_ACCESS_KEY'])
     bucket = aws_credentials[bucket]
     s3 = session.resource('s3')
     s3.meta.client.upload_file(Filename=input_path , Bucket=bucket, Key=key)
 
 def upload_pandas_to_s3(data_frame,bucket,key, aws_credentials):
+    '''
+    upload dataframe as csv to an s3 folder
 
+            Parameters:
+                    data_frame (pd.DataFrame): data
+                    bucket (str): bucket name
+                    key (str): key pattern or folder in s3 e.g. path/to/upload/
+                    aws_credentials (dict): aws credentials dictionary
+                    
+            Returns:
+                    None
+    '''
     csv_buffer = StringIO()
     data_frame.to_csv(csv_buffer)
     csv_buffer.seek(0)
@@ -23,7 +45,17 @@ def upload_pandas_to_s3(data_frame,bucket,key, aws_credentials):
     s3.put_object(Bucket=bucket, Body=csv_buffer.getvalue(), Key= key)
 
 def download_file_to_aws(bucket,key, aws_credentials):
-    
+    '''
+    download csv file from s3 folder
+
+            Parameters:
+                    bucket (str): bucket name
+                    key (str): key pattern or folder in s3 e.g. path/to/download/file.csv
+                    aws_credentials (dict): aws credentials dictionary
+                    
+            Returns:
+                    None
+    '''
     s3c = boto3.client(
             's3', 
             region_name = aws_credentials['AWS_DEFAULT_REGION'],
@@ -31,5 +63,5 @@ def download_file_to_aws(bucket,key, aws_credentials):
             aws_secret_access_key = aws_credentials['AWS_SECRET_ACCESS_KEY']
         )
     obj = s3c.get_object(Bucket= bucket , Key = key)
-    df = pd.read_csv(BytesIO(obj['Body'].read()), encoding='utf8')
+    df = pd.read_csv(BytesIO(obj['Body'].read()), encoding='utf8', sep = ';')
     return df

virgo_modules/src/backtester.py

@@ -0,0 +1,474 @@
+import pandas as pd
+import numpy as np
+import json
+
+import matplotlib.pyplot as plt
+import matplotlib.gridspec as gridspec
+import seaborn as sns; sns.set()
+
+import warnings
+warnings.filterwarnings('ignore')
+
+from .aws_utils import upload_file_to_aws
+
+def sharpe_ratio(return_series):
+
+    '''
+    calculate sharpe ratio for given array.
+
+            Parameters:
+                    return_series (pd.series): pandas series of the asset returns
+
+            Returns:
+                    sharpe (float): sharpe ratio
+    '''
+
+    N = 255 # Trading days in the year (change to 365 for crypto)
+    rf = 0.005 # Half a percent risk free rare
+    mean = return_series.mean() * N -rf
+    sigma = return_series.std() * np.sqrt(N)
+    sharpe = round(mean / sigma, 3)
+    return sharpe
+
+
+class SignalAnalyserObject:
+    """
+    Class that produces back-tests analysis for a given feature
+
+    Attributes
+    ----------
+    symbol_name : str
+        stock or asset to assess
+    feature_name : str
+    test_size: int
+        testing data size
+    show_plot: boolean
+    save_path: str
+        if available, save result locally
+    save_aws: str
+        if available, save result locally
+    aws_credentials: dict
+    signal_position: int
+        if available, signal position to open a position
+    df: pd.DataFrame
+        transformed data of the selected feature to perform back-test
+    median_return: float
+        median return after end low signals
+    
+    Methods
+    -------
+    signal_analyser(days_list=list):
+        given a signal position for either botton or roof signal, calculate the espected return and distributions for a time scope in the days list (time horizons)
+    create_backtest_signal(days_strategy=int, high_exit=float, low_exit=float, open_in_list=list):
+        create a back-test analysis using the test data using some opening anc closing postion criterias
+    """
+
+    def __init__(self, data,symbol_name, feature_name, test_size, signal_position = False, correct_signals = False, show_plot = True, save_path = False, save_aws = False, aws_credentials = False, return_fig = False):
+        """
+        Initialize object
+
+        Parameters
+        ----------
+        data (pd.DataFrame): data
+        ticket_name (str): name of the asset
+        feature_name (str): name of the features
+        test_size (int): size of the test data
+        signal_position (int): signal position to open the position, False by default
+        correct_signals (int): clean abnormal signals using interpolation
+        show_plot (boolean): if true show plot for every method
+        save_path (str): if true, save results in file e.g r'C:/path/to/the/file/'
+        save_aws (str): if true, export results to remote repo e.g. 'path/to/file/'
+        aws_credentials (dict): credentials for aws
+        return_fig (boolean): if true, methods will return objects
+
+        Returns
+        -------
+        None
+        """
+        self.ticket_name = symbol_name
+        self.feature_name=feature_name
+        self.test_size=test_size
+        self.show_plot = show_plot
+        self.save_path = save_path
+        self.save_aws = save_aws
+        self.aws_credentials = aws_credentials
+        self.return_fig = return_fig
+        self.signal_position = signal_position
+        ## preprocessing 
+        up_signal, low_signal= f'signal_up_{feature_name}', f'signal_low_{feature_name}'
+        features_base = ['Date', up_signal, low_signal, 'Close','Open','High','Low']
+
+        df = data[features_base].sort_values('Date')
+
+        df['signal_type'] = np.where(
+            df[up_signal] == 1,
+            'up',
+            np.where(
+                df[low_signal] == 1,
+                'down',
+                'no signal'
+            )
+        )
+        def correct_sygnals(df,correct_i = 1):
+            ### signal cleaning
+            for i in range(1+correct_i, len(df)-1):
+                start_i, end_i = i-(correct_i+1), i+1
+                dfw = df.iloc[start_i: end_i,]
+                before_type = dfw.iloc[0].signal_type
+                after_type = dfw.iloc[-1].signal_type
+                window_types = dfw.iloc[1:-1].signal_type.unique()
+                n_window_type = len(window_types)
+                if n_window_type == 1:
+                    if (before_type == after_type) and (window_types[0] != after_type):
+                        df.iloc[start_i+1: end_i-1, df.columns.get_loc('signal_type')] = before_type
+            return df.copy()
+
+        if correct_signals:
+            for correct_i in range(1,correct_signals+1):
+                df = correct_sygnals(df,correct_i = correct_i)
+            df[up_signal] = np.where(df['signal_type'] == 'up', 1,0)
+            df[low_signal] = np.where(df['signal_type'] == 'down', 1,0)
+
+        ## indexing chains
+        df['lag_signal_type'] = df['signal_type'].shift(1)
+        df['lag_Date'] = df['Date'].shift(1)
+        df['span'] = (pd.to_datetime(df['Date']) - pd.to_datetime(df['lag_Date'])).dt.days - 1
+        df['break'] = np.where((df['span'] > 3) & (df['lag_signal_type'] == df['signal_type']), 1, 0)
+        df['break'] = np.where((df['lag_signal_type'] != df['signal_type']), 1, df['break'])
+        df['chain_id'] = df.sort_values(['Date']).groupby(['break']).cumcount() + 1
+        df['chain_id'] = np.where(df['break'] == 1, df['chain_id'], np.nan )
+        df['chain_id'] = df['chain_id'].fillna(method = 'ffill')
+
+        df['internal_rn'] = df.sort_values(['Date']).groupby(['chain_id']).cumcount() + 1
+        df['inv_internal_rn'] = df.sort_values(['Date'],ascending = False).groupby(['chain_id']).cumcount() + 1
+
+        df['first_in_chain'] = np.where(df['internal_rn'] == 1, True, False)
+        df['last_in_chain'] = np.where(df['inv_internal_rn'] == 1, True, False)
+
+        df['span'] = (pd.to_datetime(df['Date']) - pd.to_datetime(df['lag_Date'])).dt.days - 1
+        self.df = df.drop(columns = ['span','break','lag_signal_type','lag_Date']).copy()
+        
+    def signal_analyser(self, days_list):
+        """
+        Initialize object
+
+        Parameters
+        ----------
+        days_list (list): list of integers to calculate expected returns
+
+        Returns
+        -------
+        if returns_fig is true, returns a matplotlib fig
+        """
+        signal_position = self.signal_position
+        df = self.df.iloc[0:-self.test_size,:].copy()
+        returns_list = list()
+
+        for days in days_list:
+            feature_ = f'return_{days}d'
+            df[feature_] = (df['Close'].shift(-days)/df['Close']-1)*100
+            returns_list.append(feature_)
+            
+        df['open_long'] = np.where(df.last_in_chain == True, True, np.nan)
+        df['open_short'] = np.where(df.first_in_chain == True, True, np.nan)
+        df.signal_type = df.signal_type.map({'up':'go down', 'down': 'go up'})
+        
+        # median return
+        returns_list = [f'return_{days}d' for days in days_list]
+        df_melt = df[df.open_long == True].pivot_table(index=['signal_type'], values=returns_list, aggfunc='median')
+        df_melt['median'] = df_melt[returns_list].median(axis = 1)
+        self.median_return = df_melt.loc['go up', 'median']
+        
+        # plotting
+        fig, axs = plt.subplots(1, 4, figsize = (20,5))
+        palette ={"go down": "tomato", "go up": "lightblue"}
+        
+        df2 = df[df.signal_type.isin(['go down','go up'])]
+        df2['lag_Date'] = df2['Date'].shift(1)
+        df2['lag_signal_type'] = df2['signal_type'].shift(1)
+        df2 = df2[df2.lag_signal_type != df2.signal_type]
+        df2['span'] = (pd.to_datetime(df2['Date']) - pd.to_datetime(df2['lag_Date'])).dt.days - 1
+        sns.violinplot(data=df2, y="span",ax = axs[0], color = 'lightblue', linewidth=0.7,inner="quart")
+        sns.stripplot(data=df2, y="span",ax = axs[0], jitter=True, zorder=1)
+        axs[0].set_title('span between last signals')
+
+        df_ = df[df.last_in_chain == True]
+        df_['part'] = '-'
+        sns.violinplot(data=df_, y="internal_rn", x='part', ax = axs[1], hue="signal_type", inner="quart",palette = palette,gap=0.1, split=True, linewidth=0.7)
+        axs[1].set_title('signal duration distribution')
+
+        if signal_position:
+            for feature in returns_list:
+                df[feature] = df[feature].shift(-signal_position)
+        
+        df_melt = df[df.open_long == 1].melt(id_vars=['signal_type'], value_vars=returns_list, var_name='time', value_name='value')
+        df_melt = df_melt.dropna()
+        sns.violinplot(data=df_melt, x="time", y="value", hue="signal_type",ax = axs[2], split=True, gap=0.1, inner="quart",palette = palette, linewidth=0.8)
+        axs[2].axhline(y=0, color='grey', linestyle='--')
+        axs[2].set_title('E. returns - end of the signal')
+
+        df_melt = df[df.open_short == 1].melt(id_vars=['signal_type'], value_vars=returns_list, var_name='time', value_name='value')
+        df_melt = df_melt.dropna()
+        sns.violinplot(data=df_melt, x="time", y="value", hue="signal_type",ax = axs[3], split=True, gap=0.1, inner="quart",palette = palette, linewidth=0.8)
+        axs[3].axhline(y=0, color='grey', linestyle='--')
+        axs[3].set_title('E. returns - start of the signal')
+        
+        if self.show_plot:
+            plt.show()
+
+        if self.save_path:
+            result_plot_name = f'signals_strategy_distribution_{self.feature_name}.png'
+            fig.savefig(self.save_path+result_plot_name)
+            # pickle.dump(axs, open(self.save_path+result_plot_name, 'wb'))
+
+        if self.save_path and self.save_aws:
+            # upload_file_to_aws(bucket = 'VIRGO_BUCKET', key = f'market_plots/{self.ticket_name}/'+result_plot_name, input_path = self.save_path+result_plot_name)
+            upload_file_to_aws(bucket = 'VIRGO_BUCKET', key = self.save_aws + result_plot_name, input_path = self.save_path + result_plot_name, aws_credentials = self.aws_credentials)
+        if not self.show_plot:
+            plt.close()
+
+        del df
+
+        if self.return_fig:
+            return fig
+        
+    def create_backtest_signal(self,days_strategy, high_exit = False, low_exit = False, open_in_list = ['down']):
+        """
+        Initialize object
+
+        Parameters
+        ----------
+        days_strategy (int): position horizon
+        high_exit (float): max threshold to close position
+        low_exit (float): min threshold to close position, this parameter has to be positive
+        open_in_list (list): list of strings ("down","up") to assess signals
+        Returns
+        -------
+        if returns_fig is true, returns a matplotlib fig and list of dicts containing analysis
+        """
+        asset_1 = 'Close'
+        up_signal, low_signal= f'signal_up_{self.feature_name}', f'signal_low_{self.feature_name}'
+        signal_position = self.signal_position
+        dft = self.df.iloc[-self.test_size:,:].reset_index(drop=True).copy()
+        
+        dft['lrets_bench'] = np.log(dft[asset_1]/dft[asset_1].shift(1))
+        dft['bench_prod'] = dft['lrets_bench'].cumsum()
+        dft['bench_prod_exp'] = np.exp(dft['bench_prod']) - 1
+
+        map_ = {'down':'END LOW TREND', 'up': 'BEGINNING HIGH TREND'}
+
+        open_in_list_items = len(open_in_list)
+        fig, axs = plt.subplots(1,open_in_list_items, figsize = (7*open_in_list_items,6))
+        messages = list()
+        for i, open_in in enumerate(open_in_list):
+            axs_ = axs if open_in_list_items == 1 else axs[i]
+            if open_in == 'down':
+                dft['open_long'] = np.where((dft.last_in_chain == True) & (dft.signal_type == 'down'), True, np.nan) # open strat
+            elif open_in == 'up':
+                dft['open_long'] = np.where((dft.first_in_chain == True) & (dft.signal_type == 'up'), True, np.nan) # open strat
+                
+            def chain_position(dft):
+                dft['open_long_id'] = np.where(dft['open_long'] == True, dft.chain_id, np.nan)
+                dft['open_long_id'] = dft['open_long_id'].fillna(method = 'ffill')
+                dft['open_long_rn'] = dft.sort_values(['Date']).groupby(['open_long_id']).cumcount() + 1
+                return dft
+            
+            if signal_position:
+                dft['open_long'] = dft.sort_values(['Date'])['open_long'].shift(signal_position)
+        
+            dft = chain_position(dft)
+            dft['flag'] = np.where(dft['open_long_rn'] < days_strategy, 1,0)
+        
+            if high_exit and low_exit:
+                dft['open_strat'] = np.where(dft.open_long == True, dft.Open, np.nan) # open strat
+                dft['open_strat'] = dft['open_strat'].fillna(method = 'ffill')
+                dft['open_strat'] = np.where(dft.flag == 1, dft.open_strat, np.nan)
+                dft['high_strat_ret'] = (dft['High']/dft['open_strat']-1)*100
+                dft['low_strat_ret'] = (dft['Low']/dft['open_strat']-1)*100
+                dft['max_step_chain'] = dft.groupby(['open_long_id'])['open_long_rn'].transform('max')
+                dft['high_exit'] =  np.where(((dft['high_strat_ret'] >= high_exit) | (dft['open_long_rn'] == days_strategy) | (dft['max_step_chain'] == dft['open_long_rn'])), 1, np.nan)
+                dft['low_exit'] =  np.where((dft['low_strat_ret'] <= low_exit), -1, np.nan)
+        
+                dft["exit_type"] = dft[["high_exit", "low_exit"]].max(axis=1)
+                dft['exit_type'] = np.where(dft["exit_type"] == 1, 1, np.where(dft["exit_type"] == -1,-1,np.nan))
+                dft['exit'] = np.where(dft['exit_type'].isnull(), np.nan, 1)
+                dft['exit_order'] = dft.sort_values(['Date']).groupby(['open_long_id','exit']).cumcount() + 1
+                dft['exit'] = np.where(dft['exit_order'] == 1, True, np.nan)
+                dft = dft.drop(columns = ['exit_order'])
+                ## if last signal is near
+                max_id = dft.open_long_id.max()
+                dft['max_internal_rn'] = dft.sort_values(['Date']).groupby(['open_long_id']).open_long_rn.transform('max')
+                dft['exit'] = np.where((dft.open_long_id == max_id) & (dft.max_internal_rn < days_strategy) & (dft.max_internal_rn == dft.open_long_rn), 1, dft['exit'])
+        
+                dft['exit_step'] = np.where(dft.exit == 1, dft.open_long_rn, np.nan)
+                dft['exit_step'] = dft.sort_values(['Date']).groupby(['open_long_id']).exit_step.transform('max')
+        
+                dft['flag'] = np.where(dft.open_long_rn <= dft.exit_step, 1, 0)
+        
+            dft['lrets_strat'] = np.log(dft[asset_1].shift(-1)/dft[asset_1]) * dft['flag']
+            dft['lrets_strat'] = np.where(dft['lrets_strat'].isna(),-0.0,dft['lrets_strat'])
+            dft['lrets_prod'] = dft['lrets_strat'].cumsum()
+            dft['strat_prod_exp'] = np.exp(dft['lrets_prod']) - 1
+        
+            bench_rets = round(dft['bench_prod_exp'].values[-1]*100,1)
+            strat_rets = round(dft['strat_prod_exp'].values[-1]*100,1)
+        
+            bench_sr = round(sharpe_ratio(dft.bench_prod_exp.dropna()),1)
+            strat_sr = round(sharpe_ratio(dft.strat_prod_exp.dropna()),1)
+        
+            message1 = f'{bench_rets}%'
+            message2 = f'{strat_rets}%'
+        
+            messages_ = {
+                'type strategy':map_[open_in],
+                'benchmark return:':message1,
+                'benchmark sharpe ratio:': bench_sr,
+                'strategy return:':message2,
+                'strategy sharpe ratio:': strat_sr,
+            }
+            messages.append(messages_)
+            if self.show_plot:
+                print('----------------------------')
+                print(messages_)
+                print('----------------------------')
+        
+            
+            axs_.plot(dft.bench_prod_exp.values, label = 'benchmark', color = 'steelblue')
+            axs_.scatter(range(len(dft)),np.where(dft[low_signal] == 1,dft.bench_prod_exp.values,np.nan),color = 'red', label = 'signal')
+            axs_.scatter(range(len(dft)),np.where(dft[up_signal] == 1,dft.bench_prod_exp.values,np.nan),color = 'green', label = 'signal')
+            axs_.plot(dft.strat_prod_exp.values, label = 'strategy', color = 'darksalmon')
+            axs_.set_xlabel("index")
+            axs_.set_ylabel("comulative return")
+            axs_.set_title(f'{map_[open_in]} strategy and cumulative returns')
+            axs_.legend()
+            
+        if self.show_plot:
+            plt.plot()
+
+        if self.save_path:
+            result_json_name = f'signals_strategy_return_{self.feature_name}.json'
+            result_plot_name = f'signals_strategy_return_{self.feature_name}.png'
+
+            plt.savefig(self.save_path+result_plot_name)
+
+            with open(self.save_path+result_json_name, "w") as outfile:
+                json.dump(messages, outfile)
+
+        if self.save_path and self.save_aws:
+
+            upload_file_to_aws(bucket = 'VIRGO_BUCKET', key = self.save_aws + result_json_name, input_path = self.save_path + result_json_name, aws_credentials = self.aws_credentials)
+            upload_file_to_aws(bucket = 'VIRGO_BUCKET', key = self.save_aws + result_plot_name, input_path = self.save_path + result_plot_name, aws_credentials = self.aws_credentials)
+
+        if not self.show_plot:
+            plt.close()
+
+        del dft
+
+        if self.return_fig:
+            return fig, messages
+        
+class IterateSignalAnalyse(SignalAnalyserObject):
+    """
+    object that is going to iterate backtest given a parameter space
+
+    Attributes
+    ----------
+    test_data_size : int
+    feature_name : str
+    days_list: list
+        list of integers that serve as time horizons
+    arguments_to_test : dict
+        paramter space
+    method: str
+        method to use
+    object_stock: obj
+        object containing data and methods
+    plot: boolean
+        show summary plot of median results
+    best_result: float
+        index of the best result, the index corresponds to the parameter space
+
+    Methods
+    -------
+    execute(show_plot_iter=boolean):
+        display plots for every iteration
+    """
+    def __init__(self, test_data_size, feature_name, days_list, arguments_to_test, method, object_stock, plot = False):
+        """
+        Parameters
+        ----------
+        test_data_size (int): size of the test data
+        feature_name (str): name of the feature
+        days_list (list): list of integers that serve as time horizons
+        arguments_to_test (dict): paramter space
+        method (str): method to use
+        object_stock (obj): object containing data and methods
+        plot (boolean): show summary plot of median results
+
+        Returns
+        -------
+        None
+        """
+        self.test_data_size = test_data_size
+        self.feature_name = feature_name
+        self.days_list = days_list
+        self.arguments_to_test = arguments_to_test
+        self.method = method
+        self.plot = plot
+        self.object_stock = object_stock
+        
+    def execute(self,show_plot_iter = False):
+        """
+        Iterate backtest and compute median result for every iteration
+
+        Parameters
+        ----------
+        show_plot_iter  (boolean): display plots for every iteration
+
+        Returns
+        -------
+        None
+        """
+        results = list()
+        for key in self.arguments_to_test.keys():
+            configuration = self.arguments_to_test.get(key)
+            getattr(self.object_stock, self.method)(**configuration)
+            signal_assess = SignalAnalyserObject(self.object_stock.df, self.object_stock.stock_code, show_plot = show_plot_iter, test_size = self.test_data_size, feature_name = self.feature_name)
+            signal_assess.signal_analyser(days_list = self.days_list)
+            mean_median_return = signal_assess.median_return
+            results.append(mean_median_return)
+            
+        df_result = pd.DataFrame({'keys':self.arguments_to_test.keys(),'results':results})
+        if self.plot:
+            plt.plot(df_result['keys'], df_result['results'])
+            plt.scatter(df_result['keys'], df_result['results'])
+            plt.title('simulation between configurations')
+            plt.ylabel('median expected return')
+            plt.show()
+    
+        best_result = df_result.sort_values('results',ascending = False)['keys'].values[0]
+        self.best_result = best_result
+
+def execute_signal_analyser(test_data_size, feature_name, days_list, configuration, method, object_stock, analyser_object, plot = False, backtest= False, exit_params = {}):
+    '''
+    code snippet that is going run backtest and display analysis messages and plots
+
+            Parameters:
+                    test_data_size (int): test data size
+                    feature_name (str): name of the feature to assess
+                    days_list (list): tome scope to assess the returns
+                    configuration (dict): parameters of the method to run
+                    object_stock (obj): object with data to assess
+                    method (str): method to use
+                    analyser_object (obj): signal_analyser object
+                    plot (boolean): if true, plot results
+                    backtest (boolean): if true, run backtest 
+                    exit_params (dict): parameters of exit returns
+
+            Returns:
+                    None
+    '''
+    getattr(object_stock, method)(**configuration)
+    signal_assess = analyser_object(object_stock.df,object_stock.stock_code,show_plot = plot, feature_name = feature_name, test_size = test_data_size)
+    signal_assess.signal_analyser(days_list = days_list)
+    signal_assess.create_backtest_signal(backtest, open_in_list = ['down','up'], **exit_params )

virgo_modules/src/edge_utils/conformal_utils.py

@@ -0,0 +1,106 @@
+from plotly.subplots import make_subplots
+import plotly.graph_objects as go
+from sklearn.pipeline import Pipeline
+import mlflow
+import pandas as pd
+import numpy as np
+from sklearn.base import BaseEstimator, ClassifierMixin
+from mapie.classification import SplitConformalClassifier
+
+class ConformalStack(mlflow.pyfunc.PythonModel):
+    def __init__(self, model,targets, alphas):
+        self.model = model
+        self.targets = targets
+        self.alphas = alphas
+    def fit(self, data):
+        self.classifiers = dict()
+        for i,target in enumerate(self.targets):
+            st = SingleStack(self.model["model"],i)
+            st.fit()
+            seg_model = Pipeline([ 
+                ('pipe',self.model['pipe_transform']),
+                ('modelbase',st)
+            ])
+            mapie_class = SplitConformalClassifier(seg_model, prefit=True, random_state=123, conformity_score="lac", confidence_level=1-np.array(self.alphas))
+            mapie_class.conformalize(data, data[self.targets[i]].values)
+            self.classifiers[target] = mapie_class
+    def predict_conformal(self, data, ):
+        for target in self.targets:
+            prefix = target+"_conf"
+            _, y_pis = self.classifiers[target].predict_set(data)
+            for i,alpha in enumerate(self.alphas):
+                data[f'{prefix}-{alpha}'] = y_pis[:,1,i]
+                data[f'{prefix}-{alpha}'] = np.where(data[f'{prefix}-{alpha}'] == True,alpha,0)
+        return data
+    
+
+class SingleStack(ClassifierMixin, BaseEstimator):
+    def __init__(self, model, estimator_index):
+        self.model = model
+        self.estimator_index = estimator_index
+        
+    def fit(self):
+        self._is_fitted = True
+        self.classes_ = [0,1]
+
+    def predict_proba(self, X):
+        metas_pred = dict()
+        for i,cont in enumerate(self.model.estimators, start=1):
+            _,estimator = cont
+            meta_pred = estimator.predict_proba(X)
+            metas_pred[f"meta{i}0"] = meta_pred[0][:,1]
+            metas_pred[f"meta{i}1"] = meta_pred[1][:,1]
+        self.meta_preds_df__ = pd.DataFrame(metas_pred)
+
+        prediction_vector = list()
+        for i,cont in enumerate(self.model.meta_estimators, start=0):
+            _,estimator = cont
+            metacols = [f"meta{j}{i}" for j in range(1,len(self.model.estimators)+1)]
+            preds = estimator.predict_proba(self.meta_preds_df__[metacols].values)
+            prediction_vector.append(preds)
+        return prediction_vector[self.estimator_index]
+        
+    def predict(self, X):
+        prediction_vector = list()
+        _ = self.predict_proba(X)
+        for i,cont in enumerate(self.model.meta_estimators, start=0):
+            _,estimator = cont
+            metacols = [f"meta{j}{i}" for j in range(1,len(self.model.estimators)+1)]
+            preds = estimator.predict(self.meta_preds_df__[metacols].values)
+            prediction_vector.append(preds) 
+        
+        p = np.array(tuple(prediction_vector))
+        return p.reshape((p.shape[1],p.shape[0]))[:,self.estimator_index]
+    
+    def __sklearn_is_fitted__(self):
+        return hasattr(self, "_is_fitted") and self._is_fitted
+
+def edge_conformal_lines(data, alphas,threshold = 0.6, plot = False, look_back = 750, offset = 0.08):
+    ### corect labels ####
+    df = data.sort_values('Date').iloc[-look_back:]
+    fig = make_subplots(specs=[[{"secondary_y": True}]])
+    fig.add_trace(go.Scatter(x=df.Date, y=df.Close,mode='lines+markers',marker = dict(color = 'grey'),line = dict(color = 'grey'),name='Close price'))
+    fig.add_trace(go.Scatter(x=df.Date, y=df.proba_target_up,mode='lines',marker = dict(color = 'blue'),showlegend=True,legendgroup='go up', name='go up'),secondary_y=True)
+    fig.add_trace(go.Scatter(x=df.Date, y=df.proba_target_down,mode='lines',marker = dict(color = 'coral'),showlegend=True,legendgroup='go down',name='go down'),secondary_y=True)
+    for i,alpha in enumerate(alphas, start=1):
+        try:
+            col_alpha = [x for x in df.columns if str(alpha) in x and 'target_up' in x][0]
+            df_ = df[df[col_alpha] != 0]
+            fig.add_trace(go.Scatter(x=df_.Date, y=df_.proba_target_up + (offset*i),mode='markers',marker = dict(opacity=0.7,size=10, color = 'blue')
+                                     ,showlegend=False,legendgroup='go up',name='go up', text=df_[col_alpha],textposition="bottom center")
+                                     , secondary_y=True)
+        except:
+            pass
+        try:
+            col_alpha = [x for x in df.columns if str(alpha) in x and 'target_down' in x][0]
+            df_ = df[df[col_alpha] != 0]
+            fig.add_trace(go.Scatter(x=df_.Date, y=df_.proba_target_down + (offset*i),mode='markers',marker = dict(opacity=0.7,size=10, color = 'coral')
+                                     ,showlegend=False,legendgroup='go down', name='go down',text=df_[col_alpha].astype(str),textposition="bottom center")
+                                     , secondary_y=True)
+        except:
+            pass
+    fig.add_shape(type="line", xref="paper", yref="y2",x0=0.02, y0=threshold, x1=0.9, y1=threshold,line=dict(color="red",dash="dash"))
+    fig.update_layout(title_text="sirius - edge probabilities conformal",width=1200,height = 500)
+    if plot:
+        fig.show()
+    return fig