Data-Science-Agent / src /tools /business_intelligence.py
Pulastya B
feat: Initial commit - Data Science Agent with React frontend and FastAPI backend
226ac39
"""
Business Intelligence & Analytics Tools
Advanced business analytics tools for cohort analysis, RFM segmentation,
causal inference, and automated insight generation.
"""
import polars as pl
import numpy as np
import pandas as pd
from typing import Dict, Any, List, Optional, Tuple
from datetime import datetime, timedelta
import json
# Statistical packages
try:
from scipy import stats
from scipy.stats import chi2_contingency, ttest_ind, f_oneway
except ImportError:
pass
try:
from statsmodels.tsa.stattools import grangercausalitytests
from statsmodels.stats.proportion import proportions_ztest
STATSMODELS_AVAILABLE = True
except ImportError:
STATSMODELS_AVAILABLE = False
# Causal inference (optional)
try:
from econml.dml import CausalForestDML
from econml.dr import DRLearner
ECONML_AVAILABLE = True
except ImportError:
ECONML_AVAILABLE = False
# Customer analytics (optional)
try:
from lifetimes import BetaGeoFitter, GammaGammaFitter
from lifetimes.utils import summary_data_from_transaction_data
LIFETIMES_AVAILABLE = True
except ImportError:
LIFETIMES_AVAILABLE = False
# For Groq API calls
import os
from groq import Groq
def perform_cohort_analysis(
data: pl.DataFrame,
customer_id_column: str,
date_column: str,
value_column: Optional[str] = None,
cohort_period: str = "monthly",
metric: str = "retention"
) -> Dict[str, Any]:
"""
Perform cohort analysis for customer retention, CLV, and churn analysis.
Args:
data: Input DataFrame with transaction/event data
customer_id_column: Column containing customer IDs
date_column: Column containing dates
value_column: Column containing transaction values (optional, for revenue cohorts)
cohort_period: Period for cohorts ('daily', 'weekly', 'monthly', 'quarterly')
metric: Metric to analyze ('retention', 'revenue', 'frequency', 'churn')
Returns:
Dictionary containing cohort analysis results, retention curves, and insights
"""
print(f"πŸ” Performing cohort analysis ({metric})...")
# Validate input
required_cols = [customer_id_column, date_column]
if metric == "revenue" and value_column:
required_cols.append(value_column)
for col in required_cols:
if col not in data.columns:
raise ValueError(f"Column '{col}' not found in DataFrame")
# Convert to pandas for easier date manipulation
df = data.to_pandas()
# Parse dates
df[date_column] = pd.to_datetime(df[date_column])
# Create cohort based on first purchase date
df['cohort'] = df.groupby(customer_id_column)[date_column].transform('min')
# Extract period from dates
period_map = {
'daily': 'D',
'weekly': 'W',
'monthly': 'M',
'quarterly': 'Q'
}
if cohort_period not in period_map:
raise ValueError(f"Unknown cohort_period '{cohort_period}'. Use: {list(period_map.keys())}")
period_format = {
'daily': '%Y-%m-%d',
'weekly': '%Y-W%U',
'monthly': '%Y-%m',
'quarterly': '%Y-Q%q'
}
df['cohort_period'] = df['cohort'].dt.to_period(period_map[cohort_period])
df['transaction_period'] = df[date_column].dt.to_period(period_map[cohort_period])
# Calculate period number (periods since cohort start)
df['period_number'] = (df['transaction_period'] - df['cohort_period']).apply(lambda x: x.n)
result = {
"metric": metric,
"cohort_period": cohort_period,
"total_customers": df[customer_id_column].nunique(),
"cohorts": []
}
try:
if metric == "retention":
# Retention analysis
cohort_data = df.groupby(['cohort_period', 'period_number']).agg({
customer_id_column: 'nunique'
}).reset_index()
cohort_data.columns = ['cohort_period', 'period_number', 'customers']
# Get cohort sizes (period 0)
cohort_sizes = cohort_data[cohort_data['period_number'] == 0].set_index('cohort_period')['customers']
# Calculate retention rates
cohort_data['cohort_size'] = cohort_data['cohort_period'].map(cohort_sizes)
cohort_data['retention_rate'] = cohort_data['customers'] / cohort_data['cohort_size']
# Pivot for cohort matrix
cohort_matrix = cohort_data.pivot(
index='cohort_period',
columns='period_number',
values='retention_rate'
)
result["cohort_matrix"] = cohort_matrix.to_dict()
result["avg_retention_by_period"] = cohort_matrix.mean().to_dict()
# Calculate churn (1 - retention)
result["avg_churn_by_period"] = (1 - cohort_matrix.mean()).to_dict()
# Retention curve (average across all cohorts)
retention_curve = cohort_matrix.mean().to_list()
result["retention_curve"] = retention_curve
elif metric == "revenue" and value_column:
# Revenue cohort analysis
cohort_data = df.groupby(['cohort_period', 'period_number']).agg({
value_column: 'sum',
customer_id_column: 'nunique'
}).reset_index()
cohort_data.columns = ['cohort_period', 'period_number', 'revenue', 'customers']
# Revenue per customer
cohort_data['revenue_per_customer'] = cohort_data['revenue'] / cohort_data['customers']
# Pivot for cohort matrix
cohort_matrix = cohort_data.pivot(
index='cohort_period',
columns='period_number',
values='revenue_per_customer'
)
result["cohort_matrix"] = cohort_matrix.to_dict()
result["avg_revenue_by_period"] = cohort_matrix.mean().to_dict()
# Cumulative revenue
cumulative_revenue = cohort_matrix.fillna(0).cumsum(axis=1)
result["cumulative_revenue"] = cumulative_revenue.mean().to_dict()
# Lifetime value estimate (sum of all periods)
result["estimated_ltv"] = float(cohort_matrix.sum(axis=1).mean())
elif metric == "frequency":
# Frequency analysis (purchases per period)
cohort_data = df.groupby(['cohort_period', 'period_number', customer_id_column]).size().reset_index(name='transactions')
cohort_summary = cohort_data.groupby(['cohort_period', 'period_number']).agg({
'transactions': 'mean',
customer_id_column: 'count'
}).reset_index()
cohort_summary.columns = ['cohort_period', 'period_number', 'avg_transactions', 'active_customers']
# Pivot
cohort_matrix = cohort_summary.pivot(
index='cohort_period',
columns='period_number',
values='avg_transactions'
)
result["cohort_matrix"] = cohort_matrix.to_dict()
result["avg_frequency_by_period"] = cohort_matrix.mean().to_dict()
# Cohort-level statistics
cohort_stats = []
for cohort in df['cohort_period'].unique():
cohort_df = df[df['cohort_period'] == cohort]
stats_dict = {
"cohort": str(cohort),
"size": int(cohort_df[customer_id_column].nunique()),
"total_transactions": int(len(cohort_df)),
"avg_transactions_per_customer": float(len(cohort_df) / cohort_df[customer_id_column].nunique())
}
if value_column:
stats_dict["total_revenue"] = float(cohort_df[value_column].sum())
stats_dict["avg_revenue_per_customer"] = float(cohort_df[value_column].sum() / cohort_df[customer_id_column].nunique())
cohort_stats.append(stats_dict)
result["cohort_statistics"] = cohort_stats
# Calculate key insights
result["insights"] = _generate_cohort_insights(result, metric)
print(f"βœ… Cohort analysis complete!")
print(f" Total customers: {result['total_customers']}")
print(f" Cohorts analyzed: {len(cohort_stats)}")
return result
except Exception as e:
print(f"❌ Error during cohort analysis: {str(e)}")
raise
def _generate_cohort_insights(result: Dict[str, Any], metric: str) -> List[str]:
"""Generate insights from cohort analysis."""
insights = []
if metric == "retention" and "retention_curve" in result:
retention = result["retention_curve"]
if len(retention) > 1:
initial_drop = (retention[0] - retention[1]) * 100
insights.append(f"Initial retention drop: {initial_drop:.1f}% in first period")
if len(retention) > 3:
month_3_retention = retention[3] * 100
insights.append(f"3-period retention: {month_3_retention:.1f}%")
if metric == "revenue" and "estimated_ltv" in result:
ltv = result["estimated_ltv"]
insights.append(f"Estimated customer lifetime value: ${ltv:.2f}")
return insights
def perform_rfm_analysis(
data: pl.DataFrame,
customer_id_column: str,
date_column: str,
value_column: str,
reference_date: Optional[str] = None,
rfm_bins: int = 5
) -> Dict[str, Any]:
"""
Perform RFM (Recency, Frequency, Monetary) analysis for customer segmentation.
Args:
data: Input DataFrame with transaction data
customer_id_column: Column containing customer IDs
date_column: Column containing transaction dates
value_column: Column containing transaction values
reference_date: Reference date for recency calculation (default: max date in data)
rfm_bins: Number of bins for RFM scoring (typically 3, 4, or 5)
Returns:
Dictionary containing RFM scores, segments, and customer profiles
"""
print(f"πŸ” Performing RFM analysis...")
# Validate input
required_cols = [customer_id_column, date_column, value_column]
for col in required_cols:
if col not in data.columns:
raise ValueError(f"Column '{col}' not found in DataFrame")
# Convert to pandas
df = data.to_pandas()
df[date_column] = pd.to_datetime(df[date_column])
# Set reference date
if reference_date:
ref_date = pd.to_datetime(reference_date)
else:
ref_date = df[date_column].max()
print(f" Reference date: {ref_date.strftime('%Y-%m-%d')}")
# Calculate RFM metrics
rfm = df.groupby(customer_id_column).agg({
date_column: lambda x: (ref_date - x.max()).days, # Recency
customer_id_column: 'count', # Frequency
value_column: 'sum' # Monetary
})
rfm.columns = ['recency', 'frequency', 'monetary']
# RFM Scoring (1-5, where 5 is best)
# Note: For recency, lower is better, so we reverse the scoring
rfm['r_score'] = pd.qcut(rfm['recency'], rfm_bins, labels=range(rfm_bins, 0, -1), duplicates='drop')
rfm['f_score'] = pd.qcut(rfm['frequency'].rank(method='first'), rfm_bins, labels=range(1, rfm_bins+1), duplicates='drop')
rfm['m_score'] = pd.qcut(rfm['monetary'].rank(method='first'), rfm_bins, labels=range(1, rfm_bins+1), duplicates='drop')
# Convert to int
rfm['r_score'] = rfm['r_score'].astype(int)
rfm['f_score'] = rfm['f_score'].astype(int)
rfm['m_score'] = rfm['m_score'].astype(int)
# RFM Score (concatenated)
rfm['rfm_score'] = rfm['r_score'].astype(str) + rfm['f_score'].astype(str) + rfm['m_score'].astype(str)
# RFM Total Score (sum)
rfm['rfm_total'] = rfm['r_score'] + rfm['f_score'] + rfm['m_score']
# Segment customers based on RFM scores
def segment_customer(row):
r, f, m = row['r_score'], row['f_score'], row['m_score']
if r >= 4 and f >= 4 and m >= 4:
return "Champions"
elif r >= 4 and f >= 3:
return "Loyal Customers"
elif r >= 4 and f < 3:
return "Potential Loyalists"
elif r >= 3 and f >= 3 and m >= 3:
return "Recent Customers"
elif r >= 3 and m >= 4:
return "Big Spenders"
elif r < 3 and f >= 4:
return "At Risk"
elif r < 3 and f < 3 and m >= 4:
return "Can't Lose Them"
elif r < 2:
return "Lost"
else:
return "Needs Attention"
rfm['segment'] = rfm.apply(segment_customer, axis=1)
# Results
result = {
"total_customers": len(rfm),
"reference_date": ref_date.strftime('%Y-%m-%d'),
"rfm_bins": rfm_bins,
"rfm_data": rfm.reset_index().to_dict('records'),
"segment_summary": {},
"rfm_statistics": {}
}
# Segment summary
segment_stats = rfm.groupby('segment').agg({
'recency': ['mean', 'median'],
'frequency': ['mean', 'median'],
'monetary': ['mean', 'median', 'sum'],
customer_id_column: 'count'
}).round(2)
for segment in rfm['segment'].unique():
segment_data = rfm[rfm['segment'] == segment]
result["segment_summary"][segment] = {
"count": int(len(segment_data)),
"percentage": float(len(segment_data) / len(rfm) * 100),
"avg_recency": float(segment_data['recency'].mean()),
"avg_frequency": float(segment_data['frequency'].mean()),
"avg_monetary": float(segment_data['monetary'].mean()),
"total_revenue": float(segment_data['monetary'].sum())
}
# Overall RFM statistics
result["rfm_statistics"] = {
"recency": {
"mean": float(rfm['recency'].mean()),
"median": float(rfm['recency'].median()),
"min": int(rfm['recency'].min()),
"max": int(rfm['recency'].max())
},
"frequency": {
"mean": float(rfm['frequency'].mean()),
"median": float(rfm['frequency'].median()),
"min": int(rfm['frequency'].min()),
"max": int(rfm['frequency'].max())
},
"monetary": {
"mean": float(rfm['monetary'].mean()),
"median": float(rfm['monetary'].median()),
"min": float(rfm['monetary'].min()),
"max": float(rfm['monetary'].max()),
"total": float(rfm['monetary'].sum())
}
}
# Top customers by RFM score
result["top_customers"] = rfm.nlargest(20, 'rfm_total').reset_index().to_dict('records')
# Actionable insights
result["recommendations"] = _generate_rfm_recommendations(result)
print(f"βœ… RFM analysis complete!")
print(f" Total customers: {result['total_customers']}")
print(f" Segments: {len(result['segment_summary'])}")
print(f" Top segment: {max(result['segment_summary'].items(), key=lambda x: x[1]['count'])[0]}")
return result
def _generate_rfm_recommendations(result: Dict[str, Any]) -> Dict[str, List[str]]:
"""Generate actionable recommendations based on RFM segments."""
recommendations = {}
segment_actions = {
"Champions": [
"Reward with exclusive perks and early access to new products",
"Request reviews and referrals",
"Engage for product development feedback"
],
"Loyal Customers": [
"Upsell higher value products",
"Offer loyalty rewards",
"Encourage referrals with incentives"
],
"Potential Loyalists": [
"Recommend related products",
"Offer membership or loyalty program",
"Engage with personalized communication"
],
"Recent Customers": [
"Provide onboarding support",
"Build relationships with targeted content",
"Offer starter discounts for repeat purchases"
],
"Big Spenders": [
"Target with premium products",
"Increase engagement frequency",
"Offer VIP treatment"
],
"At Risk": [
"Send win-back campaigns",
"Offer special discounts or incentives",
"Gather feedback on their experience"
],
"Can't Lose Them": [
"Aggressive win-back campaigns",
"Personalized outreach",
"Offer significant incentives"
],
"Lost": [
"Run re-engagement campaigns",
"Survey for feedback",
"Consider removing from active campaigns"
],
"Needs Attention": [
"Offer limited-time promotions",
"Share valuable content",
"Re-engage with surveys"
]
}
for segment, actions in segment_actions.items():
if segment in result["segment_summary"]:
recommendations[segment] = actions
return recommendations
def detect_causal_relationships(
data: pl.DataFrame,
treatment_column: str,
outcome_column: str,
covariates: Optional[List[str]] = None,
method: str = "granger",
max_lag: int = 5,
confidence_level: float = 0.95
) -> Dict[str, Any]:
"""
Detect causal relationships using Granger causality, propensity matching, or uplift modeling.
Args:
data: Input DataFrame
treatment_column: Column indicating treatment/intervention
outcome_column: Column indicating outcome variable
covariates: List of covariate columns for adjustment
method: Method for causal inference ('granger', 'propensity', 'uplift')
max_lag: Maximum lag for Granger causality test
confidence_level: Confidence level for statistical tests
Returns:
Dictionary containing causal inference results and effect estimates
"""
print(f"πŸ” Detecting causal relationships using {method} method...")
# Validate input
required_cols = [treatment_column, outcome_column]
if covariates:
required_cols.extend(covariates)
for col in required_cols:
if col not in data.columns:
raise ValueError(f"Column '{col}' not found in DataFrame")
result = {
"method": method,
"treatment": treatment_column,
"outcome": outcome_column,
"covariates": covariates or [],
"causal_effect": None,
"statistical_significance": None
}
try:
if method == "granger" and STATSMODELS_AVAILABLE:
# Granger causality test for time series
print(f" Testing Granger causality with max lag = {max_lag}...")
# Convert to pandas
df = data.select([treatment_column, outcome_column]).to_pandas()
# Ensure numeric
df = df.apply(pd.to_numeric, errors='coerce').dropna()
# Test both directions
test_result = grangercausalitytests(
df[[outcome_column, treatment_column]],
max_lag,
verbose=False
)
# Extract p-values for each lag
granger_results = []
for lag in range(1, max_lag + 1):
ssr_ftest = test_result[lag][0]['ssr_ftest']
granger_results.append({
"lag": lag,
"f_statistic": float(ssr_ftest[0]),
"p_value": float(ssr_ftest[1]),
"significant": ssr_ftest[1] < (1 - confidence_level)
})
result["granger_causality"] = granger_results
result["causal_effect"] = any(r["significant"] for r in granger_results)
result["statistical_significance"] = min(r["p_value"] for r in granger_results)
elif method == "propensity":
# Propensity score matching
print(" Performing propensity score matching...")
df = data.to_pandas()
# Ensure treatment is binary
treatment = df[treatment_column]
if treatment.nunique() > 2:
raise ValueError(f"Treatment column must be binary for propensity matching")
outcome = df[outcome_column]
# Simple comparison without covariates
if not covariates:
treated = outcome[treatment == 1]
control = outcome[treatment == 0]
# Calculate average treatment effect
ate = treated.mean() - control.mean()
# T-test for significance
t_stat, p_value = ttest_ind(treated, control)
result["average_treatment_effect"] = float(ate)
result["t_statistic"] = float(t_stat)
result["p_value"] = float(p_value)
result["statistical_significance"] = float(p_value)
result["causal_effect"] = float(ate)
result["confidence_interval"] = [
float(ate - 1.96 * np.sqrt(treated.var()/len(treated) + control.var()/len(control))),
float(ate + 1.96 * np.sqrt(treated.var()/len(treated) + control.var()/len(control)))
]
else:
# With covariates (simplified - use logistic regression for propensity)
from sklearn.linear_model import LogisticRegression
from sklearn.neighbors import NearestNeighbors
X = df[covariates].apply(pd.to_numeric, errors='coerce').fillna(0)
# Estimate propensity scores
ps_model = LogisticRegression(max_iter=1000)
ps_model.fit(X, treatment)
propensity_scores = ps_model.predict_proba(X)[:, 1]
df['propensity_score'] = propensity_scores
# Matching (1:1 nearest neighbor)
treated_df = df[treatment == 1]
control_df = df[treatment == 0]
# Simple matching on propensity scores
nn = NearestNeighbors(n_neighbors=1)
nn.fit(control_df[['propensity_score']])
distances, indices = nn.kneighbors(treated_df[['propensity_score']])
matched_control = control_df.iloc[indices.flatten()]
# Calculate ATE on matched sample
ate = treated_df[outcome_column].mean() - matched_control[outcome_column].mean()
result["average_treatment_effect"] = float(ate)
result["n_matched_pairs"] = len(treated_df)
result["causal_effect"] = float(ate)
elif method == "uplift":
# Uplift modeling (treatment effect heterogeneity)
print(" Calculating uplift/treatment effect...")
df = data.to_pandas()
treatment = df[treatment_column]
outcome = df[outcome_column]
# Calculate uplift by treatment group
treated_outcome = outcome[treatment == 1].mean()
control_outcome = outcome[treatment == 0].mean()
uplift = treated_outcome - control_outcome
# Statistical significance
t_stat, p_value = ttest_ind(
outcome[treatment == 1],
outcome[treatment == 0]
)
result["uplift"] = float(uplift)
result["treated_mean"] = float(treated_outcome)
result["control_mean"] = float(control_outcome)
result["relative_uplift"] = float(uplift / control_outcome * 100) if control_outcome != 0 else 0
result["t_statistic"] = float(t_stat)
result["p_value"] = float(p_value)
result["statistical_significance"] = float(p_value)
result["causal_effect"] = float(uplift)
else:
raise ValueError(f"Unknown method '{method}'. Use 'granger', 'propensity', or 'uplift'")
print(f"βœ… Causal analysis complete!")
if result.get("causal_effect") is not None:
print(f" Estimated causal effect: {result['causal_effect']:.4f}")
return result
except Exception as e:
print(f"❌ Error during causal analysis: {str(e)}")
raise
def generate_business_insights(
data: pl.DataFrame,
analysis_type: str,
analysis_results: Dict[str, Any],
additional_context: Optional[str] = None,
groq_api_key: Optional[str] = None
) -> Dict[str, Any]:
"""
Generate natural language business insights using Groq LLM.
Args:
data: Input DataFrame (for context)
analysis_type: Type of analysis ('rfm', 'cohort', 'causal', 'general')
analysis_results: Results from previous analysis (dict)
additional_context: Additional business context
groq_api_key: Groq API key (if not in environment)
Returns:
Dictionary containing natural language insights and recommendations
"""
print(f"πŸ” Generating business insights for {analysis_type} analysis...")
# Get API key
api_key = groq_api_key or os.getenv("GROQ_API_KEY")
if not api_key:
raise ValueError("Groq API key not found. Set GROQ_API_KEY environment variable or pass groq_api_key parameter")
client = Groq(api_key=api_key)
# Prepare data summary
data_summary = {
"shape": data.shape,
"columns": data.columns,
"dtypes": {col: str(dtype) for col, dtype in zip(data.columns, data.dtypes)},
"sample_stats": {}
}
# Add numeric column stats
for col in data.columns:
if data[col].dtype in [pl.Int32, pl.Int64, pl.Float32, pl.Float64]:
data_summary["sample_stats"][col] = {
"mean": float(data[col].mean()),
"median": float(data[col].median()),
"std": float(data[col].std()),
"min": float(data[col].min()),
"max": float(data[col].max())
}
# Create prompt based on analysis type
prompt = f"""You are a senior business analyst. Analyze the following data and provide actionable business insights.
Analysis Type: {analysis_type.upper()}
Data Summary:
{json.dumps(data_summary, indent=2)}
Analysis Results:
{json.dumps(analysis_results, indent=2)}
Additional Context:
{additional_context or 'None provided'}
Please provide:
1. Key findings (3-5 bullet points)
2. Business implications
3. Actionable recommendations (3-5 specific actions)
4. Risk factors or caveats
5. Suggested next steps
Format your response as a structured business report."""
try:
# Call Groq API
response = client.chat.completions.create(
model="llama-3.3-70b-versatile",
messages=[
{
"role": "system",
"content": "You are a senior business analyst specializing in data-driven insights and strategic recommendations. Provide clear, actionable insights based on data analysis."
},
{
"role": "user",
"content": prompt
}
],
temperature=0.3,
max_tokens=2000
)
insights_text = response.choices[0].message.content
# Parse insights (simple structure)
result = {
"analysis_type": analysis_type,
"insights_summary": insights_text,
"generated_at": datetime.now().isoformat(),
"model": "llama-3.3-70b-versatile",
"data_context": data_summary
}
# Try to extract structured sections
sections = {}
current_section = None
for line in insights_text.split('\n'):
line = line.strip()
if line.startswith('1.') or line.lower().startswith('key findings'):
current_section = 'key_findings'
sections[current_section] = []
elif line.startswith('2.') or line.lower().startswith('business implications'):
current_section = 'implications'
sections[current_section] = []
elif line.startswith('3.') or line.lower().startswith('actionable recommendations'):
current_section = 'recommendations'
sections[current_section] = []
elif line.startswith('4.') or line.lower().startswith('risk'):
current_section = 'risks'
sections[current_section] = []
elif line.startswith('5.') or line.lower().startswith('next steps'):
current_section = 'next_steps'
sections[current_section] = []
elif current_section and line:
sections[current_section].append(line)
result["structured_insights"] = sections
print(f"βœ… Business insights generated!")
print(f" Sections: {', '.join(sections.keys())}")
return result
except Exception as e:
print(f"❌ Error generating insights: {str(e)}")
raise