app.py

# Import necessary libraries
import streamlit as st
import pandas as pd
import numpy as np
from plotly import graph_objs as go
import joblib
import cloudpickle

from xgboost import XGBRegressor
from sklearn.preprocessing import StandardScaler
from sklearn.preprocessing import MinMaxScaler
from sklearn.preprocessing import RobustScaler

from skforecast.utils import save_forecaster
from skforecast.utils import load_forecaster
from skforecast.ForecasterAutoreg import ForecasterAutoreg

from sklearn.metrics import mean_squared_error, mean_absolute_percentage_error

# ========================================== Helper Functions ==========================================

def evaluate_forecast(y_true, y_pred):
    results = {
        'RMSE': np.sqrt(mean_squared_error(y_true, y_pred)),
        'MAPE': mean_absolute_percentage_error(y_true, y_pred)
    }
    return pd.Series(results)

# Define functions for transformations
def apply_transformation(data, transform_type):
    if transform_type == 'Log':
        return np.log1p(data)
    elif transform_type == 'Square Root':
        return np.sqrt(data)
    else:
        return data

def reverse_transformation(transformed_data, transform_type):
    if transform_type == 'Log':
        return np.expm1(transformed_data)
    elif transform_type == 'Square Root':
        return np.square(transformed_data)
    else:
        return transformed_data

# Cached function for auto-tuning
@st.cache_data
def run_auto_tuning(train, test, lags_to_try, differentiation_options, transformer_options, external_transform_options):
    results = []
    for lag in lags_to_try:
        for diff in differentiation_options:
            for trans in transformer_options:
                for ext_trans in external_transform_options:
                    # Apply External Transformation
                    train_transformed = apply_transformation(train[target_column], ext_trans)

                    # Transformer Selection
                    transformer_y = select_transformer(trans)

                    # Create and fit the forecaster
                    forecaster = ForecasterAutoreg(
                                    regressor = XGBRegressor(random_state=123),
                                    lags = lag,
                                    differentiation = diff,
                                    transformer_y = transformer_y
                                )
                    forecaster.fit(y=train_transformed)

                    # Predictions and Evaluation
                    predictions = forecaster.predict(steps=len(test))
                    predictions_reversed = reverse_transformation(predictions, ext_trans)
                    actual = test[target_column].iloc[:len(predictions)]
                    rmse = np.sqrt(mean_squared_error(actual, predictions_reversed))
                    mape = mean_absolute_percentage_error(actual, predictions_reversed)

                    # Store results
                    results.append({
                        'Lag': lag,
                        'Differentiation': diff,
                        'Transformer': trans,
                        'External Transformer': ext_trans,
                        'RMSE': rmse,
                        'MAPE': mape
                    })

    return pd.DataFrame(results)

# Helper function to select transformer
def select_transformer(transformer_option):
    if transformer_option == 'StandardScaler':
        return StandardScaler()
    elif transformer_option == 'MinMaxScaler':
        return MinMaxScaler()
    elif transformer_option == 'RobustScaler':
        return RobustScaler()
    return None

@st.cache_resource
def train_model(lags,differentiation,_transformer_y,train_data):

    # Create and fit forecaster

    forecaster = \
        ForecasterAutoreg(regressor=XGBRegressor(random_state=123),
                          lags=lags, differentiation=differentiation,
                          transformer_y=transformer_y)

    forecaster.fit(y=train_data)
    save_forecaster(forecaster, file_name='forecaster_temp.py',
                    verbose=False)
    return forecaster

@st.cache_data
def predict(_forecaster, n_steps, external_transform, test, target_column):
    predictions = forecaster.predict(steps=n_steps)
    predictions_reversed = reverse_transformation(predictions, external_transform)

    # Prepare Comparison DataFrame
    actual = test[target_column].iloc[:len(predictions)]
    pred = predictions_reversed.to_frame(name='Predicted')
    comparison_df = pd.concat([actual.reset_index(drop=True), pred.reset_index(drop=True)], axis=1)
    evaluation_results = evaluate_forecast(comparison_df[target_column], comparison_df['Predicted'])

    return predictions_reversed, actual, pred, comparison_df, evaluation_results

# Function to load and cache the data
@st.cache_data
def load_data(uploaded_file):
    return pd.read_excel(uploaded_file)

@st.cache_resource
def refit(_forecaster, df, target_column, external_transform):
    entire_data_transformed = apply_transformation(df[target_column], external_transform)
    forecaster.fit(y=entire_data_transformed)
    return forecaster

# ========================================== Header ==========================================

# Streamlit app layout
st.title("SKForecast Forecasting App")
st.write("Upload an xlsx file for time series analysis")

# ========================================== Section: Load Data ==========================================
st.header("Load Data")
uploaded_file = st.file_uploader("Choose a file", type="xlsx")

if uploaded_file is not None:
    # Load and cache the dataframe
    df = load_data(uploaded_file)

    st.write("Dataframe:")
    st.write(df)

    # ========================================== Section: Select Data ==========================================
    st.header("Select Data")
    date_column = st.selectbox("Select Date Column", df.columns)
    target_column = st.selectbox("Select Target Column", [col for col in df.columns if col != date_column])

    if date_column != target_column:
        df[date_column] = pd.to_datetime(df[date_column])
        df.set_index(date_column, inplace=True)

				# Date Range Selection
        st.subheader("Filter Date Range")
        start_date = st.date_input("Start Date", value=df.index.min(), min_value=df.index.min(), max_value=df.index.max())
        end_date = st.date_input("End Date", value=df.index.max(), min_value=df.index.min(), max_value=df.index.max())
        df = df[start_date:end_date]

        freq_option = st.selectbox("Select Frequency for Resampling", ['No Resampling', 'W-SUN', 'W-MON', 'W-TUE', 'W-WED', 'W-THU', 'W-FRI', 'W-SAT', 'M', 'MS'])
        if freq_option != 'No Resampling':
            df = df.resample(freq_option).mean()

        st.write("Selected Data with Datetime Index:")
        st.write(df[[target_column]])

        # ========================================== Section: Split Data ==========================================
        st.header("Split Data")
        split_method = st.radio("Select Method for Train-Test Split", ('Percentage', 'Size', 'Year Range', 'Specific Year'))

        if split_method == 'Percentage':
            split_type = st.radio("Select Split Type", ('Training Set', 'Testing Set'))
            if split_type == 'Training Set':
                percentage = st.slider("Select Percentage for Training Set", 0.1, 0.85, 0.7)
                split_point = int(len(df) * percentage)
            else:
                percentage = st.slider("Select Percentage for Testing Set", 0.15, 0.9, 0.15)
                split_point = int(len(df) * (1 - percentage))
            train = df.iloc[:split_point]
            test = df.iloc[split_point:]

        elif split_method == 'Size':
            split_type = st.radio("Select Split Type", ('Training Set', 'Testing Set'))
            max_train_size = int(0.9 * len(df))
            max_test_size = int(0.9 * len(df))
            if split_type == 'Training Set':
                size = st.number_input("Enter Size for Training Set", 1, max_train_size, max_train_size)
                train = df.iloc[:size]
                test = df.iloc[size:]
            else:
                size = st.number_input("Enter Size for Testing Set", 1, max_test_size, max_test_size)
                train = df.iloc[:-size]
                test = df.iloc[-size:]

        elif split_method == 'Year Range':
            start_year = st.selectbox("Select Start Year", range(df.index.year.min(), df.index.year.max() + 1))
            end_year = st.selectbox("Select End Year", range(start_year, df.index.year.max() + 1))
            train = df[(df.index.year >= start_year) & (df.index.year <= end_year)]
            test = df.drop(train.index)

        elif split_method == 'Specific Year':
            split_type = st.radio("Select Split Type", ('Training Set', 'Testing Set'))
            year = st.selectbox("Select Year", range(df.index.year.min(), df.index.year.max() + 1))
            if split_type == 'Training Set':
                train = df[df.index.year <= year]
                test = df[df.index.year > year]
            else:
                test = df[df.index.year == year]
                train = df.drop(test.index)

        # ========================================== Section: Display Sets and Visualize ==========================================
        st.header("Display Data and Visualize Split")
        col1, col2 = st.columns(2)
        with col1:
            st.write("Training Set:")
            st.write(train[target_column])
        with col2:
            st.write("Test Set:")
            st.write(test[target_column])

        # Plotting both Sets
        fig = go.Figure()
        fig.add_trace(go.Scatter(x=train.index, y=train[target_column], mode='lines', name='Training Set', line=dict(color='aqua')))
        fig.add_trace(go.Scatter(x=test.index, y=test[target_column], mode='lines', name='Test Set', line=dict(color='orange')))
        fig.update_layout(title='Train-Test Split Visualization', xaxis_title='Date', yaxis_title=target_column)
        st.plotly_chart(fig)


        # Initialize session state for auto-tuning results
        if 'auto_tuning_results' not in st.session_state:
            st.session_state.auto_tuning_results = None

        # ========================================== Section: Auto-Tuning ==========================================
        st.header("Auto-Tuning")
        st.write("Automatically test various configurations to identify the optimal setup")

        # Input for Lag Ranges
        lag_input = st.text_input("Enter Lag Ranges (e.g. 1,2,3-5)", "1,2,3-5")

        # Parsing lag ranges
        lags_to_try = []
        for part in lag_input.split(','):
            if '-' in part:
                a, b = part.split('-')
                lags_to_try.extend(range(int(a), int(b) + 1))
            else:
                lags_to_try.append(int(part))

        # Other Parameters
        differentiation_options = [None, 1, 2]
        transformer_options = [None, 'StandardScaler', 'MinMaxScaler', 'RobustScaler']
        external_transform_options = [None, 'Log', 'Square Root']

        # Run Button for Auto-Tuning
        if st.button("Run Auto-Tuning"):
            st.cache_data.clear()
            # Run the cached auto-tuning function
            auto_tuning_results = run_auto_tuning(train, test, lags_to_try, differentiation_options, transformer_options, external_transform_options)

            # Storing best configurations in session state
            st.session_state.best_config_rmse = auto_tuning_results.sort_values(by='RMSE').iloc[0]
            st.session_state.best_config_mape = auto_tuning_results.sort_values(by='MAPE').iloc[0]

            st.session_state.auto_tuning_results = auto_tuning_results
            st.success("Auto-tuning finished!")

        # Display auto-tuning results from session state
        if st.session_state.auto_tuning_results is not None:
            st.write("Auto-Tuning Results:")
            st.write(st.session_state.auto_tuning_results.sort_values(by='MAPE'))

            # Display Best Configurations for Each Metric
            col1, col2 = st.columns(2)
            with col1:
                st.write("Best Configuration for RMSE:", st.session_state.best_config_rmse)
            with col2:
                st.write("Best Configuration for MAPE:", st.session_state.best_config_mape)

        # ========================================== Section: Train Model ==========================================
        st.header("Train Model")
        
        # Initialize session state for prediction results
        if 'forecaster' not in st.session_state:
            st.session_state.forecaster = None
            st.session_state.final_forecaster = None

        if 'train' in locals():
            # Check if auto-tuning results are available and valid
            if ('auto_tuning_results' in st.session_state and
                    isinstance(st.session_state.auto_tuning_results, pd.DataFrame) and
                    not st.session_state.auto_tuning_results.empty):

                auto_tuned_config_option = st.radio(
                    "Choose Configuration to Use",
                    ('Manual Configuration', 'Best RMSE Configuration', 'Best MAPE Configuration')
                )

                if auto_tuned_config_option != 'Manual Configuration':
                    if auto_tuned_config_option == 'Best RMSE Configuration':
                        best_config = st.session_state.auto_tuning_results.sort_values(by='RMSE').iloc[0]
                    elif auto_tuned_config_option == 'Best MAPE Configuration':
                        best_config = st.session_state.auto_tuning_results.sort_values(by='MAPE').iloc[0]

                    lags = int(best_config['Lag'])  # Convert to regular Python integer
                    differentiation = int(best_config['Differentiation']) if pd.notna(best_config['Differentiation']) else None
                    transformer_y = best_config['Transformer']
                    external_transform = best_config['External Transformer']
                else:
                    # Manual configuration
                    lags = st.slider("Select Lags", 1, max(1, int(len(train) * 0.5)), 4)
                    differentiation = st.selectbox("Select Differentiation Order", [None, 1, 2])
                    transformer_y = st.selectbox("Select Transformer", [None, 'StandardScaler', 'MinMaxScaler', 'RobustScaler'])
                    external_transform = st.selectbox("Select External Transformation", [None, 'Log', 'Square Root'])

            else:
                # Only manual configuration available
                st.write("Manual Configuration:")
                lags = st.slider("Select Lags", 1, max(1, int(len(train) * 0.5)), 4)
                differentiation = st.selectbox("Select Differentiation Order", [None, 1, 2])
                transformer_y = st.selectbox("Select Transformer", [None, 'StandardScaler', 'MinMaxScaler', 'RobustScaler'])
                external_transform = st.selectbox("Select External Transformation", [None, 'Log', 'Square Root'])

            # Apply External Transformation
            train_transformed = apply_transformation(train[target_column], external_transform)

            # Train Button
            if st.button("Train"):
                st.cache_resource.clear()
                with st.spinner('Training in progress...'):
                    if transformer_y == 'StandardScaler':
                        transformer_y = StandardScaler()
                    elif transformer_y == 'MinMaxScaler':
                        transformer_y = MinMaxScaler()
                    elif transformer_y == 'RobustScaler':
                        transformer_y = RobustScaler()
                    else:
                        transformer_y = None

                    forecaster = train_model(lags, differentiation, transformer_y, train_transformed)
                    save_forecaster(forecaster, file_name='forecaster_temp.py', verbose=False)
                    st.session_state.forecaster = forecaster
                    st.success("Model trained successfully!")
        else:
            st.warning("Please complete the 'Split Data' section first.")
        # ========================================== Section: Predict ==========================================

        # Initialize session state for prediction results
        if 'comparison_df' not in st.session_state:
            st.session_state.comparison_df = None
            st.session_state.predictions_reversed = None
            st.session_state.pred = None
            st.session_state.actual = None
            st.session_state.evaluation_results = None

        st.header("Predict")
        st.subheader("Forecast Configuration")
        default_steps = len(test) if 'test' in locals() else 1
        n_steps = st.number_input("Number of Steps for Prediction", 1, len(df), default_steps)

        # Predict Button
        if st.button("Predict"):
            st.cache_data.clear()
            forecaster = st.session_state.forecaster
            st.session_state.predictions_reversed, st.session_state.actual, st.session_state.pred, st.session_state.comparison_df, st.session_state.evaluation_results = predict(forecaster, n_steps, external_transform, test, target_column)

        if st.session_state.comparison_df is not None:
            # Display Predictions vs Actual
            st.subheader("Predictions vs Actual Values")
            st.write(st.session_state.comparison_df)

            # Plotting Predictions vs Actual
            fig = go.Figure()
            fig.add_trace(go.Scatter(y=st.session_state.actual, mode='lines', name='Actual'))
            fig.add_trace(go.Scatter(y=st.session_state.pred['Predicted'], mode='lines', name='Predicted'))
            fig.update_layout(title='Actual vs Predicted Values', xaxis_title='Index', yaxis_title=target_column)
            st.plotly_chart(fig)

            # Plotting Train + Actual vs Train + Predicted
            fig_comparison = go.Figure()
            fig_comparison.add_trace(go.Scatter(x=train.index, y=train[target_column], mode='lines', name='Train'))
            fig_comparison.add_trace(go.Scatter(x=st.session_state.actual.index, y=st.session_state.actual, mode='lines', name='Actual'))
            fig_comparison.add_trace(go.Scatter(x=st.session_state.pred.index, y=st.session_state.pred['Predicted'], mode='lines', name='Predicted'))
            fig_comparison.update_layout(title='Train, Actual vs Predicted Values', xaxis_title='Date', yaxis_title=target_column)
            st.plotly_chart(fig_comparison)

            # Enhanced Evaluation Results Display
            st.subheader("Model Evaluation Results")

            col1, col2 = st.columns(2)
            with col1:
                st.metric(label="RMSE", value=f"{st.session_state.evaluation_results['RMSE']:.3f}")
            with col2:
                st.metric(label="MAPE", value=f"{st.session_state.evaluation_results['MAPE']*100:.3f} %")

        # ========================================== Section: Save & Download Model ==========================================
        st.header("Save & Download Model")

        # Refit Model
        if st.button("Refit Model on Entire Dataset"):
            forecaster = st.session_state.forecaster
            st.session_state.final_forecaster = refit(forecaster, df, target_column, external_transform)
            st.success("Model refitted on the entire dataset.")
        else:
            st.session_state.final_forecaster = st.session_state.forecaster

        save_method = st.selectbox("Select Save Method", ['SKForecast', 'Joblib', 'Pickle'])
        model_name = st.text_input("Enter Model Name", 'forecaster_model')

        # Save/Download Button
        if save_method == 'SKForecast':
            file_name = f'{model_name}.py'
            save_forecaster(st.session_state.final_forecaster, file_name=file_name, verbose=False)
            st.download_button(label="Download Model as SKForecast", data=open(file_name, "rb").read(), file_name=file_name, mime='text/plain')

        elif save_method == 'Joblib':
            file_name = f'{model_name}.joblib'
            joblib.dump(st.session_state.final_forecaster, filename=file_name)
            st.download_button(label="Download Model as Joblib", data=open(file_name, "rb").read(), file_name=file_name, mime='application/octet-stream')

        elif save_method == 'Pickle':
            file_name = f'{model_name}.pkl'
            with open(file_name, 'wb') as file:
                cloudpickle.dump(st.session_state.final_forecaster, file)
            st.download_button(label="Download Model as Pickle", data=open(file_name, "rb").read(), file_name=file_name, mime='application/octet-stream')

    else:
        st.error("Date column and Target column cannot be the same. Please select different columns.")
else:
    st.warning("Please upload an xlsx file to proceed.")