""" Economic Forecasting Module Advanced time series forecasting for economic indicators using ARIMA/ETS models """ import logging import warnings from datetime import datetime, timedelta from typing import Dict, List, Optional, Tuple, Union import numpy as np import pandas as pd from scipy import stats from sklearn.metrics import mean_absolute_error, mean_squared_error from statsmodels.tsa.arima.model import ARIMA from statsmodels.tsa.holtwinters import ExponentialSmoothing from statsmodels.tsa.seasonal import seasonal_decompose from statsmodels.tsa.stattools import adfuller logger = logging.getLogger(__name__) class EconomicForecaster: """ Advanced economic forecasting using ARIMA and ETS models with comprehensive backtesting and performance evaluation """ def __init__(self, data: pd.DataFrame): """ Initialize forecaster with economic data Args: data: DataFrame with economic indicators (GDPC1, INDPRO, RSAFS, etc.) """ self.data = data.copy() self.forecasts = {} self.backtest_results = {} self.model_performance = {} def prepare_data(self, target_series: str, frequency: str = 'Q', for_arima: bool = True) -> pd.Series: """ Prepare time series data for forecasting or analysis. Args: target_series: Series name to forecast frequency: Data frequency ('Q' for quarterly, 'M' for monthly) for_arima: If True, returns raw levels for ARIMA; if False, returns growth rate Returns: Prepared time series """ if target_series not in self.data.columns: raise ValueError(f"Series {target_series} not found in data") series = self.data[target_series].dropna() # Ensure time-based index if not isinstance(series.index, pd.DatetimeIndex): raise ValueError("Index must be datetime type") # Resample to desired frequency if frequency == 'Q': series = series.resample('Q').mean() elif frequency == 'M': series = series.resample('M').mean() # Only use growth rates if for_arima is False if not for_arima and target_series in ['GDPC1', 'INDPRO', 'RSAFS']: series = series.pct_change().dropna() return series def check_stationarity(self, series: pd.Series) -> Dict: """ Perform Augmented Dickey-Fuller test for stationarity Args: series: Time series to test Returns: Dictionary with test results """ result = adfuller(series.dropna()) return { 'adf_statistic': result[0], 'p_value': result[1], 'critical_values': result[4], 'is_stationary': result[1] < 0.05 } def decompose_series(self, series: pd.Series, period: int = 4) -> Dict: """ Decompose time series into trend, seasonal, and residual components Args: series: Time series to decompose period: Seasonal period (4 for quarterly, 12 for monthly) Returns: Dictionary with decomposition components """ decomposition = seasonal_decompose(series.dropna(), period=period, extrapolate_trend='freq') return { 'trend': decomposition.trend, 'seasonal': decomposition.seasonal, 'residual': decomposition.resid, 'observed': decomposition.observed } def fit_arima_model(self, series: pd.Series, order: Tuple[int, int, int] = None) -> ARIMA: """ Fit ARIMA model to time series using raw levels (not growth rates) Args: series: Time series data (raw levels) order: ARIMA order (p, d, q). If None, auto-detect Returns: Fitted ARIMA model """ # Ensure we're working with raw levels, not growth rates if series.isna().any(): series = series.dropna() # Ensure series has enough data points if len(series) < 10: raise ValueError("Series must have at least 10 data points for ARIMA fitting") if order is None: # Auto-detect order using AIC minimization with improved search best_aic = np.inf best_order = (1, 1, 1) # Improved order search that avoids degenerate models # Start with more reasonable orders to avoid ARIMA(0,0,0) search_orders = [ (1, 1, 1), (2, 1, 1), (1, 1, 2), (2, 1, 2), # Common orders (0, 1, 1), (1, 0, 1), (1, 1, 0), # Simple orders (2, 0, 1), (1, 0, 2), (2, 1, 0), # Alternative orders (3, 1, 1), (1, 1, 3), (2, 2, 1), (1, 2, 2), # Higher orders ] for p, d, q in search_orders: try: model = ARIMA(series, order=(p, d, q)) fitted_model = model.fit() # Check if model is degenerate (all parameters near zero) params = fitted_model.params if len(params) > 0: # Skip models where all AR/MA parameters are very small ar_params = params[1:p+1] if p > 0 else [] ma_params = params[p+1:p+1+q] if q > 0 else [] # Check if model is essentially a random walk or constant if (p == 0 and d == 0 and q == 0) or \ (p == 0 and d == 1 and q == 0) or \ (len(ar_params) > 0 and all(abs(p) < 0.01 for p in ar_params)) or \ (len(ma_params) > 0 and all(abs(p) < 0.01 for p in ma_params)): logger.debug(f"Skipping degenerate ARIMA({p},{d},{q})") continue if fitted_model.aic < best_aic: best_aic = fitted_model.aic best_order = (p, d, q) logger.debug(f"New best ARIMA({p},{d},{q}) with AIC: {best_aic}") except Exception as e: logger.debug(f"ARIMA({p},{d},{q}) failed: {e}") continue order = best_order logger.info(f"Auto-detected ARIMA order: {order} with AIC: {best_aic}") # If we still have a degenerate model, force a reasonable order if order == (0, 0, 0) or order == (0, 1, 0): logger.warning("Detected degenerate ARIMA order, forcing to ARIMA(1,1,1)") order = (1, 1, 1) try: model = ARIMA(series, order=order) fitted_model = model.fit() # Debug: Log model parameters logger.info(f"ARIMA model fitted successfully with AIC: {fitted_model.aic}") logger.info(f"ARIMA order: {order}") logger.info(f"Model parameters: {fitted_model.params}") return fitted_model except Exception as e: logger.warning(f"ARIMA fitting failed with order {order}: {e}") # Try fallback orders fallback_orders = [(1, 1, 1), (0, 1, 1), (1, 0, 1), (1, 1, 0)] for fallback_order in fallback_orders: try: model = ARIMA(series, order=fallback_order) fitted_model = model.fit() logger.info(f"ARIMA fallback model fitted with order {fallback_order}") return fitted_model except Exception as fallback_e: logger.debug(f"Fallback ARIMA{fallback_order} failed: {fallback_e}") continue # Last resort: simple moving average logger.warning("All ARIMA models failed, using simple moving average") raise ValueError("Unable to fit any ARIMA model to the data") def fit_ets_model(self, series: pd.Series, seasonal_periods: int = 4) -> ExponentialSmoothing: """ Fit ETS (Exponential Smoothing) model to time series Args: series: Time series data seasonal_periods: Number of seasonal periods Returns: Fitted ETS model """ model = ExponentialSmoothing( series, seasonal_periods=seasonal_periods, trend='add', seasonal='add' ) fitted_model = model.fit() return fitted_model def forecast_series(self, series: pd.Series, model_type: str = 'auto', forecast_periods: int = 4) -> Dict: """ Forecast time series using specified model Args: series: Time series to forecast model_type: 'arima', 'ets', or 'auto' forecast_periods: Number of periods to forecast Returns: Dictionary with forecast results """ if model_type == 'auto': # Try both models and select the one with better AIC try: arima_model = self.fit_arima_model(series) arima_aic = arima_model.aic except: arima_aic = np.inf try: ets_model = self.fit_ets_model(series) ets_aic = ets_model.aic except: ets_aic = np.inf if arima_aic < ets_aic: model_type = 'arima' model = arima_model else: model_type = 'ets' model = ets_model elif model_type == 'arima': model = self.fit_arima_model(series) elif model_type == 'ets': model = self.fit_ets_model(series) else: raise ValueError("model_type must be 'arima', 'ets', or 'auto'") # Generate forecast using proper method for each model type if model_type == 'arima': # Use get_forecast() for ARIMA to get proper confidence intervals forecast_result = model.get_forecast(steps=forecast_periods) forecast = forecast_result.predicted_mean try: forecast_ci = forecast_result.conf_int() # Check if confidence intervals are valid (not all NaN) if forecast_ci.isna().all().all() or forecast_ci.empty: # Improved fallback confidence intervals forecast_ci = self._calculate_improved_confidence_intervals(forecast, series, model) else: # Ensure confidence intervals have proper column names if len(forecast_ci.columns) >= 2: forecast_ci.columns = ['lower', 'upper'] else: # Improved fallback if column structure is unexpected forecast_ci = self._calculate_improved_confidence_intervals(forecast, series, model) # Debug: Log confidence intervals logger.info(f"ARIMA confidence intervals: {forecast_ci.to_dict()}") # Check if confidence intervals are too wide and provide warning ci_widths = forecast_ci['upper'] - forecast_ci['lower'] mean_width = ci_widths.mean() forecast_mean = forecast.mean() relative_width = mean_width / abs(forecast_mean) if abs(forecast_mean) > 0 else 0 if relative_width > 0.5: # If confidence interval is more than 50% of forecast value logger.warning(f"Confidence intervals are very wide (relative width: {relative_width:.2%})") logger.info("This may indicate high uncertainty or model instability") except Exception as e: logger.warning(f"ARIMA confidence interval calculation failed: {e}") # Improved fallback confidence intervals forecast_ci = self._calculate_improved_confidence_intervals(forecast, series, model) else: # For ETS, use forecast() method forecast = model.forecast(steps=forecast_periods) # Use improved confidence intervals for ETS forecast_ci = self._calculate_improved_confidence_intervals(forecast, series, model) # Debug: Log final results logger.info(f"Final forecast is flat: {len(set(forecast)) == 1}") logger.info(f"Forecast type: {type(forecast)}") return { 'model': model, 'model_type': model_type, 'forecast': forecast, 'confidence_intervals': forecast_ci, 'aic': model.aic if hasattr(model, 'aic') else None } def _calculate_improved_confidence_intervals(self, forecast: pd.Series, series: pd.Series, model) -> pd.DataFrame: """ Calculate improved confidence intervals with better uncertainty quantification Args: forecast: Forecast values series: Original time series model: Fitted model Returns: DataFrame with improved confidence intervals """ try: # Calculate forecast errors from model residuals if available if hasattr(model, 'resid') and len(model.resid) > 0: # Use model residuals for more accurate uncertainty residuals = model.resid.dropna() forecast_std = residuals.std() # Adjust for forecast horizon (uncertainty increases with horizon) horizon_factors = np.sqrt(np.arange(1, len(forecast) + 1)) confidence_intervals = [] for i, (fcast, factor) in enumerate(zip(forecast, horizon_factors)): # Use 95% confidence interval (1.96 * std) margin = 1.96 * forecast_std * factor lower = fcast - margin upper = fcast + margin confidence_intervals.append({'lower': lower, 'upper': upper}) return pd.DataFrame(confidence_intervals, index=forecast.index) else: # Fallback to series-based uncertainty series_std = series.std() # Use a more conservative approach for economic data # Economic forecasts typically have higher uncertainty uncertainty_factor = 1.5 # Adjust based on data characteristics confidence_intervals = [] for i, fcast in enumerate(forecast): # Increase uncertainty with forecast horizon horizon_factor = 1 + (i * 0.1) # 10% increase per period margin = 1.96 * series_std * uncertainty_factor * horizon_factor lower = fcast - margin upper = fcast + margin confidence_intervals.append({'lower': lower, 'upper': upper}) return pd.DataFrame(confidence_intervals, index=forecast.index) except Exception as e: logger.warning(f"Improved confidence interval calculation failed: {e}") # Ultimate fallback series_std = series.std() return pd.DataFrame({ 'lower': forecast - 1.96 * series_std, 'upper': forecast + 1.96 * series_std }, index=forecast.index) def backtest_forecast(self, series: pd.Series, model_type: str = 'auto', train_size: float = 0.8, test_periods: int = 8) -> Dict: """ Perform backtesting of forecasting models Args: series: Time series to backtest model_type: Model type to use train_size: Proportion of data for training test_periods: Number of periods to test Returns: Dictionary with backtest results """ n = len(series) train_end = int(n * train_size) actual_values = [] predicted_values = [] errors = [] for i in range(test_periods): if train_end + i >= n: break # Use expanding window train_data = series.iloc[:train_end + i] test_value = series.iloc[train_end + i] try: forecast_result = self.forecast_series(train_data, model_type, 1) prediction = forecast_result['forecast'].iloc[0] actual_values.append(test_value) predicted_values.append(prediction) errors.append(test_value - prediction) except Exception as e: logger.warning(f"Forecast failed at step {i}: {e}") continue if not actual_values: return {'error': 'No successful forecasts generated'} # Calculate performance metrics mae = mean_absolute_error(actual_values, predicted_values) mse = mean_squared_error(actual_values, predicted_values) rmse = np.sqrt(mse) # Use safe MAPE calculation to avoid division by zero actual_array = np.array(actual_values) predicted_array = np.array(predicted_values) denominator = np.maximum(np.abs(actual_array), 1e-8) mape = np.mean(np.abs((actual_array - predicted_array) / denominator)) * 100 return { 'actual_values': actual_values, 'predicted_values': predicted_values, 'errors': errors, 'mae': mae, 'mse': mse, 'rmse': rmse, 'mape': mape, 'test_periods': len(actual_values) } def forecast_economic_indicators(self, indicators: List[str] = None) -> Dict: """ Forecast multiple economic indicators Args: indicators: List of indicators to forecast. If None, use default set Returns: Dictionary with forecasts for all indicators """ if indicators is None: indicators = ['GDPC1', 'INDPRO', 'RSAFS'] results = {} for indicator in indicators: try: # Prepare raw data for forecasting (use raw levels, not growth rates) series = self.prepare_data(indicator, for_arima=True) # Prepare growth rates for analysis growth_series = self.prepare_data(indicator, for_arima=False) # Check stationarity on growth rates stationarity = self.check_stationarity(growth_series) # Decompose growth rates decomposition = self.decompose_series(growth_series) # Generate forecast using raw levels forecast_result = self.forecast_series(series) # Perform backtest on raw levels backtest_result = self.backtest_forecast(series) results[indicator] = { 'stationarity': stationarity, 'decomposition': decomposition, 'forecast': forecast_result, 'backtest': backtest_result, 'raw_series': series, 'growth_series': growth_series } logger.info(f"Successfully forecasted {indicator}") except Exception as e: logger.error(f"Failed to forecast {indicator}: {e}") results[indicator] = {'error': str(e)} return results def generate_forecast_report(self, forecast_result, periods=None): """ Generate a markdown table for forecast results. Args: forecast_result: dict with keys 'forecast', 'confidence_intervals' periods: list of period labels (optional) Returns: Markdown string """ forecast = forecast_result.get('forecast') ci = forecast_result.get('confidence_intervals') if forecast is None or ci is None: return 'No forecast results available.' if periods is None: periods = [f"Period {i+1}" for i in range(len(forecast))] lines = ["| Period | Forecast | 95% CI Lower | 95% CI Upper |", "| ------- | ------------- | ------------ | ------------ |"] for i, (f, p) in enumerate(zip(forecast, periods)): try: lower = ci.iloc[i, 0] if hasattr(ci, 'iloc') else ci[i][0] upper = ci.iloc[i, 1] if hasattr(ci, 'iloc') else ci[i][1] except Exception: lower = upper = 'N/A' lines.append(f"| {p} | **{f:,.2f}** | {lower if isinstance(lower, str) else f'{lower:,.2f}'} | {upper if isinstance(upper, str) else f'{upper:,.2f}'} |") return '\n'.join(lines)