FREDML / src /analysis /statistical_modeling.py
Edwin Salguero
feat: Integrate advanced analytics and enterprise UI
26a8ea5
raw
history blame
20.1 kB
"""
Statistical Modeling Module
Advanced statistical analysis for economic indicators including regression, correlation, and diagnostics
"""
import logging
from typing import Dict, List, Optional, Tuple, Union
import numpy as np
import pandas as pd
from scipy import stats
from sklearn.linear_model import LinearRegression
from sklearn.metrics import r2_score, mean_squared_error
from sklearn.preprocessing import StandardScaler
from statsmodels.stats.diagnostic import het_breuschpagan
from statsmodels.stats.outliers_influence import variance_inflation_factor
from statsmodels.stats.stattools import durbin_watson
from statsmodels.tsa.stattools import adfuller, kpss
logger = logging.getLogger(__name__)
class StatisticalModeling:
"""
Advanced statistical modeling for economic indicators
including regression analysis, correlation analysis, and diagnostic testing
"""
def __init__(self, data: pd.DataFrame):
"""
Initialize statistical modeling with economic data
Args:
data: DataFrame with economic indicators
"""
self.data = data.copy()
self.models = {}
self.diagnostics = {}
self.correlations = {}
def prepare_regression_data(self, target: str, predictors: List[str] = None,
lag_periods: int = 4) -> Tuple[pd.DataFrame, pd.Series]:
"""
Prepare data for regression analysis with lagged variables
Args:
target: Target variable name
predictors: List of predictor variables. If None, use all other numeric columns
lag_periods: Number of lag periods to include
Returns:
Tuple of (features DataFrame, target Series)
"""
if target not in self.data.columns:
raise ValueError(f"Target variable {target} not found in data")
if predictors is None:
predictors = [col for col in self.data.select_dtypes(include=[np.number]).columns
if col != target]
# Calculate growth rates for all variables
growth_data = self.data[[target] + predictors].pct_change().dropna()
# Create lagged features
feature_data = {}
for predictor in predictors:
# Current value
feature_data[predictor] = growth_data[predictor]
# Lagged values
for lag in range(1, lag_periods + 1):
feature_data[f"{predictor}_lag{lag}"] = growth_data[predictor].shift(lag)
# Add target variable lags as features
for lag in range(1, lag_periods + 1):
feature_data[f"{target}_lag{lag}"] = growth_data[target].shift(lag)
# Create feature matrix
features_df = pd.DataFrame(feature_data)
features_df = features_df.dropna()
# Target variable
target_series = growth_data[target].iloc[features_df.index]
return features_df, target_series
def fit_regression_model(self, target: str, predictors: List[str] = None,
lag_periods: int = 4, include_interactions: bool = False) -> Dict:
"""
Fit linear regression model with diagnostic testing
Args:
target: Target variable name
predictors: List of predictor variables
lag_periods: Number of lag periods to include
include_interactions: Whether to include interaction terms
Returns:
Dictionary with model results and diagnostics
"""
# Prepare data
features_df, target_series = self.prepare_regression_data(target, predictors, lag_periods)
if include_interactions:
# Add interaction terms
interaction_features = []
feature_cols = features_df.columns.tolist()
for i, col1 in enumerate(feature_cols):
for col2 in feature_cols[i+1:]:
interaction_name = f"{col1}_x_{col2}"
interaction_features.append(features_df[col1] * features_df[col2])
features_df[interaction_name] = interaction_features[-1]
# Scale features
scaler = StandardScaler()
features_scaled = scaler.fit_transform(features_df)
features_scaled_df = pd.DataFrame(features_scaled,
index=features_df.index,
columns=features_df.columns)
# Fit model
model = LinearRegression()
model.fit(features_scaled_df, target_series)
# Predictions
predictions = model.predict(features_scaled_df)
residuals = target_series - predictions
# Model performance
r2 = r2_score(target_series, predictions)
mse = mean_squared_error(target_series, predictions)
rmse = np.sqrt(mse)
# Coefficient analysis
coefficients = pd.DataFrame({
'variable': features_df.columns,
'coefficient': model.coef_,
'abs_coefficient': np.abs(model.coef_)
}).sort_values('abs_coefficient', ascending=False)
# Diagnostic tests
diagnostics = self.perform_regression_diagnostics(features_scaled_df, target_series,
predictions, residuals)
return {
'model': model,
'scaler': scaler,
'features': features_df,
'target': target_series,
'predictions': predictions,
'residuals': residuals,
'coefficients': coefficients,
'performance': {
'r2': r2,
'mse': mse,
'rmse': rmse,
'mae': np.mean(np.abs(residuals))
},
'diagnostics': diagnostics
}
def perform_regression_diagnostics(self, features: pd.DataFrame, target: pd.Series,
predictions: np.ndarray, residuals: pd.Series) -> Dict:
"""
Perform comprehensive regression diagnostics
Args:
features: Feature matrix
target: Target variable
predictions: Model predictions
residuals: Model residuals
Returns:
Dictionary with diagnostic test results
"""
diagnostics = {}
# 1. Normality test (Shapiro-Wilk)
try:
normality_stat, normality_p = stats.shapiro(residuals)
diagnostics['normality'] = {
'statistic': normality_stat,
'p_value': normality_p,
'is_normal': normality_p > 0.05
}
except:
diagnostics['normality'] = {'error': 'Test failed'}
# 2. Homoscedasticity test (Breusch-Pagan)
try:
bp_stat, bp_p, bp_f, bp_f_p = het_breuschpagan(residuals, features)
diagnostics['homoscedasticity'] = {
'statistic': bp_stat,
'p_value': bp_p,
'f_statistic': bp_f,
'f_p_value': bp_f_p,
'is_homoscedastic': bp_p > 0.05
}
except:
diagnostics['homoscedasticity'] = {'error': 'Test failed'}
# 3. Autocorrelation test (Durbin-Watson)
try:
dw_stat = durbin_watson(residuals)
diagnostics['autocorrelation'] = {
'statistic': dw_stat,
'interpretation': self._interpret_durbin_watson(dw_stat)
}
except:
diagnostics['autocorrelation'] = {'error': 'Test failed'}
# 4. Multicollinearity test (VIF)
try:
vif_scores = {}
for i, col in enumerate(features.columns):
vif = variance_inflation_factor(features.values, i)
vif_scores[col] = vif
diagnostics['multicollinearity'] = {
'vif_scores': vif_scores,
'high_vif_variables': [var for var, vif in vif_scores.items() if vif > 10],
'mean_vif': np.mean(list(vif_scores.values()))
}
except:
diagnostics['multicollinearity'] = {'error': 'Test failed'}
# 5. Stationarity tests
try:
# ADF test
adf_result = adfuller(target)
diagnostics['stationarity_adf'] = {
'statistic': adf_result[0],
'p_value': adf_result[1],
'is_stationary': adf_result[1] < 0.05
}
# KPSS test
kpss_result = kpss(target, regression='c')
diagnostics['stationarity_kpss'] = {
'statistic': kpss_result[0],
'p_value': kpss_result[1],
'is_stationary': kpss_result[1] > 0.05
}
except:
diagnostics['stationarity'] = {'error': 'Test failed'}
return diagnostics
def _interpret_durbin_watson(self, dw_stat: float) -> str:
"""Interpret Durbin-Watson statistic"""
if dw_stat < 1.5:
return "Positive autocorrelation"
elif dw_stat > 2.5:
return "Negative autocorrelation"
else:
return "No significant autocorrelation"
def analyze_correlations(self, indicators: List[str] = None,
method: str = 'pearson') -> Dict:
"""
Perform comprehensive correlation analysis
Args:
indicators: List of indicators to analyze. If None, use all numeric columns
method: Correlation method ('pearson', 'spearman', 'kendall')
Returns:
Dictionary with correlation analysis results
"""
if indicators is None:
indicators = self.data.select_dtypes(include=[np.number]).columns.tolist()
# Calculate growth rates
growth_data = self.data[indicators].pct_change().dropna()
# Correlation matrix
corr_matrix = growth_data.corr(method=method)
# Significant correlations
significant_correlations = []
for i in range(len(corr_matrix.columns)):
for j in range(i+1, len(corr_matrix.columns)):
var1 = corr_matrix.columns[i]
var2 = corr_matrix.columns[j]
corr_value = corr_matrix.iloc[i, j]
# Test significance
n = len(growth_data)
t_stat = corr_value * np.sqrt((n-2) / (1-corr_value**2))
p_value = 2 * (1 - stats.t.cdf(abs(t_stat), n-2))
if p_value < 0.05:
significant_correlations.append({
'variable1': var1,
'variable2': var2,
'correlation': corr_value,
'p_value': p_value,
'strength': self._interpret_correlation_strength(abs(corr_value))
})
# Sort by absolute correlation
significant_correlations.sort(key=lambda x: abs(x['correlation']), reverse=True)
# Principal Component Analysis
try:
pca = self._perform_pca_analysis(growth_data)
except Exception as e:
logger.warning(f"PCA analysis failed: {e}")
pca = {'error': str(e)}
return {
'correlation_matrix': corr_matrix,
'significant_correlations': significant_correlations,
'method': method,
'pca_analysis': pca
}
def _interpret_correlation_strength(self, corr_value: float) -> str:
"""Interpret correlation strength"""
if corr_value >= 0.8:
return "Very Strong"
elif corr_value >= 0.6:
return "Strong"
elif corr_value >= 0.4:
return "Moderate"
elif corr_value >= 0.2:
return "Weak"
else:
return "Very Weak"
def _perform_pca_analysis(self, data: pd.DataFrame) -> Dict:
"""Perform Principal Component Analysis"""
from sklearn.decomposition import PCA
# Standardize data
scaler = StandardScaler()
data_scaled = scaler.fit_transform(data)
# Perform PCA
pca = PCA()
pca_result = pca.fit_transform(data_scaled)
# Explained variance
explained_variance = pca.explained_variance_ratio_
cumulative_variance = np.cumsum(explained_variance)
# Component loadings
loadings = pd.DataFrame(
pca.components_.T,
columns=[f'PC{i+1}' for i in range(pca.n_components_)],
index=data.columns
)
return {
'explained_variance': explained_variance,
'cumulative_variance': cumulative_variance,
'loadings': loadings,
'n_components': pca.n_components_,
'components_to_explain_80_percent': np.argmax(cumulative_variance >= 0.8) + 1
}
def perform_granger_causality(self, target: str, predictor: str,
max_lags: int = 4) -> Dict:
"""
Perform Granger causality test
Args:
target: Target variable
predictor: Predictor variable
max_lags: Maximum number of lags to test
Returns:
Dictionary with Granger causality test results
"""
try:
from statsmodels.tsa.stattools import grangercausalitytests
# Prepare data
growth_data = self.data[[target, predictor]].pct_change().dropna()
# Perform Granger causality test
test_data = growth_data[[predictor, target]] # Note: order matters
gc_result = grangercausalitytests(test_data, maxlag=max_lags, verbose=False)
# Extract results
results = {}
for lag in range(1, max_lags + 1):
if lag in gc_result:
lag_result = gc_result[lag]
results[lag] = {
'f_statistic': lag_result[0]['ssr_ftest'][0],
'p_value': lag_result[0]['ssr_ftest'][1],
'is_significant': lag_result[0]['ssr_ftest'][1] < 0.05
}
# Overall result (use minimum p-value)
min_p_value = min([result['p_value'] for result in results.values()])
overall_significant = min_p_value < 0.05
return {
'results_by_lag': results,
'min_p_value': min_p_value,
'is_causal': overall_significant,
'optimal_lag': min(results.keys(), key=lambda k: results[k]['p_value'])
}
except Exception as e:
logger.error(f"Granger causality test failed: {e}")
return {'error': str(e)}
def generate_statistical_report(self, regression_results: Dict = None,
correlation_results: Dict = None,
causality_results: Dict = None) -> str:
"""
Generate comprehensive statistical analysis report
Args:
regression_results: Results from regression analysis
correlation_results: Results from correlation analysis
causality_results: Results from causality analysis
Returns:
Formatted report string
"""
report = "STATISTICAL MODELING REPORT\n"
report += "=" * 50 + "\n\n"
if regression_results:
report += "REGRESSION ANALYSIS\n"
report += "-" * 30 + "\n"
# Model performance
performance = regression_results['performance']
report += f"Model Performance:\n"
report += f" R²: {performance['r2']:.4f}\n"
report += f" RMSE: {performance['rmse']:.4f}\n"
report += f" MAE: {performance['mae']:.4f}\n\n"
# Top coefficients
coefficients = regression_results['coefficients']
report += f"Top 5 Most Important Variables:\n"
for i, row in coefficients.head().iterrows():
report += f" {row['variable']}: {row['coefficient']:.4f}\n"
report += "\n"
# Diagnostics
diagnostics = regression_results['diagnostics']
report += f"Model Diagnostics:\n"
if 'normality' in diagnostics and 'error' not in diagnostics['normality']:
norm = diagnostics['normality']
report += f" Normality (Shapiro-Wilk): p={norm['p_value']:.4f} "
report += f"({'Normal' if norm['is_normal'] else 'Not Normal'})\n"
if 'homoscedasticity' in diagnostics and 'error' not in diagnostics['homoscedasticity']:
hom = diagnostics['homoscedasticity']
report += f" Homoscedasticity (Breusch-Pagan): p={hom['p_value']:.4f} "
report += f"({'Homoscedastic' if hom['is_homoscedastic'] else 'Heteroscedastic'})\n"
if 'autocorrelation' in diagnostics and 'error' not in diagnostics['autocorrelation']:
autocorr = diagnostics['autocorrelation']
report += f" Autocorrelation (Durbin-Watson): {autocorr['statistic']:.4f} "
report += f"({autocorr['interpretation']})\n"
if 'multicollinearity' in diagnostics and 'error' not in diagnostics['multicollinearity']:
mult = diagnostics['multicollinearity']
report += f" Multicollinearity (VIF): Mean VIF = {mult['mean_vif']:.2f}\n"
if mult['high_vif_variables']:
report += f" High VIF variables: {', '.join(mult['high_vif_variables'])}\n"
report += "\n"
if correlation_results:
report += "CORRELATION ANALYSIS\n"
report += "-" * 30 + "\n"
report += f"Method: {correlation_results['method'].title()}\n"
report += f"Significant Correlations: {len(correlation_results['significant_correlations'])}\n\n"
# Top correlations
report += f"Top 5 Strongest Correlations:\n"
for i, corr in enumerate(correlation_results['significant_correlations'][:5]):
report += f" {corr['variable1']}{corr['variable2']}: "
report += f"{corr['correlation']:.4f} ({corr['strength']}, p={corr['p_value']:.4f})\n"
# PCA results
if 'pca_analysis' in correlation_results and 'error' not in correlation_results['pca_analysis']:
pca = correlation_results['pca_analysis']
report += f"\nPrincipal Component Analysis:\n"
report += f" Components to explain 80% variance: {pca['components_to_explain_80_percent']}\n"
report += f" Total components: {pca['n_components']}\n"
report += "\n"
if causality_results:
report += "GRANGER CAUSALITY ANALYSIS\n"
report += "-" * 30 + "\n"
for target, results in causality_results.items():
if 'error' not in results:
report += f"{target}:\n"
report += f" Is causal: {results['is_causal']}\n"
report += f" Minimum p-value: {results['min_p_value']:.4f}\n"
report += f" Optimal lag: {results['optimal_lag']}\n\n"
return report