actuaflow

ActuaFlow

Modern Actuarial Pricing Library for Non-Life Insurance

ActuaFlow is a Python library for building GLM-based frequency-severity pricing models with a focus on actuarial best practices.

Features

GLM Framework

• Distribution families: Poisson, Negative Binomial, Gamma, Inverse Gaussian, Lognormal, Tweedie
• Link functions: Log, identity, inverse, inverse squared
• Proper offset handling for exposure-based frequency models
• Comprehensive diagnostics: AIC, BIC, deviance, dispersion, overdispersion tests
• Model validation: VIF (variance inflation factors), correlation analysis
• Predictive performance: Lift curves, Gini index, residual plots
• Formula parsing: R-style formula notation with interactions and transformations
• Robust input validation: NaN detection, response/weight validation, formula parsing
• GAM: actuaflow.glm.GAM for smooth non-linear predictor effects
• MonotoneGLM: actuaflow.glm.MonotoneGLM applies post-fit monotonicity constraints (increasing/decreasing) to ordered categorical predictors
• QuantileGLM: actuaflow.glm.QuantileGLM conditional quantile regression (statsmodels QuantReg) with severity fan plots
• BayesianGLM: actuaflow.glm.BayesianGLM PyMC-based Bayesian Poisson/Gamma GLM with credible intervals (requires optional pymc dependency)
• HierarchicalGLM: actuaflow.glm.HierarchicalGLM mixed-effects/hierarchical GLM (statsmodels MixedLM, with optional pymer4/rpy2 support for non-Gaussian families)

Frequency-Severity Workflow

• FrequencyModel: Poisson/Negative Binomial models for claim counts
• SeverityModel: Gamma/Lognormal/Inverse Gaussian for claim amounts
• AggregateModel: Combined frequency-severity with pure premium calculation
• Automatic data preparation: Policy+claims aggregation, automatic frequency table generation
• Flexible column naming: Support for different policy/claim ID column names
• Frequency-Severity dependence modeling: FrequencySeverityCopula for joint frequency/severity simulation, tail dependence, and aggregate loss analysis
• Quantile Premium Principle (QPP): actuaflow.glm.qpp (FrequencyModel, QPPAdjuster, QPPPricer) prices policies from the policy-specific severity quantile implied by the frequency model's zero-claim probability, rather than the expected loss
• Premium loading: Sequential loadings for inflation, expenses, commission, profit, taxes
• Export capabilities: Model summaries, factor relativities, rating tables

Exposure & Rating Tools

• Exposure calculation: Days, months, years methods with custom date columns
• Rate per exposure: Calculate rate per unit exposure given pure premium
• Class plan creation: Rating tables with base rates and factor relativities
• Factor application: Apply factor levels from rating tables to policies
• Trend adjustment: Historical-to-current trending with configurable rates
• On-leveling: Mid-term rate change adjustments
• Experience modification: Experience rating factors (EMR/XMod)
• Credibility weighting: Complement experience data with prior information
• Tariff plans: actuaflow.tariff (TariffPlan, FactorTable, RatingEngine) for base rates, factor relativities, tier adjustments, eligibility rules, JSON/YAML serialization, and version comparison; build_tariff_from_glm() builds a tariff plan directly from a fitted GLM's coefficients

Portfolio & Impact Analysis

• Premium impact analysis: Calculate impact of factor changes on portfolio premium
• Mix-shift decomposition: Separate impact of rate changes from volume/mix changes
• Segment impact: Winners and losers analysis by segment
• Rate adequacy: Loss ratio, earned premium, change metrics
• Price elasticity: Demand curve estimation and revenue optimization
• Elasticity curves: Revenue impact curves across price ranges
• Retention modeling: Retention rate impact on portfolio profitability
• Scenario analysis: Compare multiple rating strategies
• Constrained rate optimization: ConstrainedRateOptimizer combines credibility-bounded and linear-programming rate optimization with dislocation analysis and dislocation plots

Regulatory Compliance

• Rate indication summaries: actuaflow.compliance.RateIndicationSummary reports actual loss ratios, indicated rates, credibility weights, and selection flags for rate review
• Rate filing packages: actuaflow.compliance.RateFilingPackage bundles rate indications and dislocation analyses, and runs anti-discrimination checks on protected-class attributes
• SERFF export: actuaflow.compliance.SERFFExporter exports rating relativities to SERFF-compliant XML or flat CSV for regulatory submission (e.g. AR, CA, GR filings)

Data Utilities

• Data loading: CSV/Parquet support with polars fast reader
• Data validation: Schema checking, missing value detection, type validation
• Data quality checks: DataQualityChecker for missing-value/outlier reports, winsorising, factor consolidation, auto-binning, leakage checks, and calibration curves
• Auto-imputation: AutoImputer for basic missing-value imputation
• Train-test split: Stratified splits with random state control
• Time series cross-validation: Temporal splits for time-dependent data
• Cross-validation scoring: Model performance across folds

Advanced Features

• Statsmodels integration: Leverages industry-standard GLM library
• Robust error handling: Custom exceptions with detailed diagnostics
• Model serialization: Export/import for production deployment
• Testing: 68%+ test coverage with pytest
• Documentation: Docstrings, examples, user guide

Current release notes

• Package version: 0.3.1
• New in this version: Model event bus (ModelFittedEvent, ModelSavedEvent), fitting strategy pattern (FittingStrategy), and ModelStore persistence.
• Behavior note: SeverityGLM.fit performs zero filtering automatically and lognormal family is internally handled with transformation.
• New FrequencySeverityCopula workflow for joint frequency/severity dependence modeling and aggregate loss simulation.
• New GLM variants: GAM, MonotoneGLM, QuantileGLM, BayesianGLM (optional pymc dependency), HierarchicalGLM (mixed effects).
• New Quantile Premium Principle (QPP) pricing workflow in actuaflow.glm.qpp (FrequencyModel, QPPAdjuster, QPPPricer). Note: actuaflow.glm.qpp.FrequencyModel is a lightweight QPP-specific wrapper and is distinct from the top-level actuaflow.FrequencyModel frequency-severity workflow class.
• New tariff/rating engine (actuaflow.tariff: TariffPlan, FactorTable, RatingEngine) and constrained rate optimizer (ConstrainedRateOptimizer) with dislocation analysis.
• New regulatory compliance module (actuaflow.compliance: RateIndicationSummary, RateFilingPackage, SERFFExporter) for rate filing and SERFF export.
• New data quality tools (DataQualityChecker, AutoImputer) in actuaflow.data.

Installation

pip install actuaflow

Or install from source:

git clone https://github.com/WattyAI/actuaflow.git
cd actuaflow
pip install -e .

Quick Start

Basic Frequency-Severity Model

import actuaflow as af
import pandas as pd

# Load data
policies = af.load_data('policies.csv')
claims = af.load_data('claims.csv')

# 1. Frequency Model
freq_model = af.FrequencyModel(family='poisson', link='log')
freq_model.prepare_data(policies, claims, policy_id='policy_id')
freq_model.fit(
    formula='claim_count ~ age_group + vehicle_type + region',
    offset='exposure'
)

print(freq_model.summary())
print(f"Dispersion: {freq_model.diagnostics_['dispersion']:.3f}")

# 2. Severity Model
sev_model = af.SeverityModel(family='gamma', link='log')
sev_model.prepare_data(claims, policies, policy_id='policy_id')
sev_model.fit(formula='amount ~ age_group + injury_type')

print(sev_model.summary())

# 3. Combined Model
agg_model = af.combine_models(freq_model, sev_model)
print(f"Base Pure Premium: ${agg_model.base_pure_premium_:.2f}")

# Get factor table
factor_table = agg_model.create_factor_table()
print(factor_table)

# 4. Premium Calculation
pure_premium = agg_model.predict_pure_premium(policies, exposure='exposure')

loadings = {
    'inflation': 0.03,
    'expense_ratio': 0.15,
    'commission': 0.10,
    'profit_margin': 0.05,
    'tax_rate': 0.02
}

premium_df = af.calculate_premium(pure_premium, loadings, exposure=policies['exposure'])
print(f"Total Loaded Premium: ${premium_df['loaded_premium'].sum():,.0f}")

Model Diagnostics

Working with Different Column Names

ActuaFlow functions that prepare frequency data accept flexible column names for policy and claims identifiers. Example:

from actuaflow import FrequencyModel

# Policy table uses 'policy_key', claims table uses 'pol_id'
freq_model = FrequencyModel()
freq_model.prepare_data(
    policy_data=policies,
    claims_data=claims,
    policy_id_policy='policy_key',
    policy_id_claims='pol_id'
)

This makes it easy to work with data exports that use different naming conventions.

Calculating Policy Exposure

Use the calculate_exposure helper to compute exposure between a start and end date. It supports years, days, and months methods and accepts custom column names for the start/end dates as well as a custom output column name.

from actuaflow import calculate_exposure

policies_with_exposure = calculate_exposure(
    policy_data=policies,
    start_date_col='policy_start_date',
    end_date_col='policy_end_date',
    method='years',
    exposure_col='exposure'
)

print(policies_with_exposure[['policy_id', 'exposure']].head())

The function validates date columns and returns a copy of the dataframe with the exposure column added.

# Comprehensive diagnostics
diag = af.compute_diagnostics(freq_model.model_, policies)

print(f"AIC: {diag['aic']:.2f}")
print(f"Dispersion: {diag['dispersion']:.3f}")

# VIF for multicollinearity
if diag['vif']:
    for var, vif in diag['vif'].items():
        print(f"{var}: VIF = {vif:.2f}")

# Lift curve
y_true = policies['claim_count']
y_pred = freq_model.predict()
bins, actual, predicted = af.compute_lift_curve(y_true, y_pred, n_bins=10)

# Gini index
has_claim = (y_true > 0).astype(int)
gini = af.compute_gini_index(has_claim, y_pred)
print(f"Gini coefficient: {gini:.3f}")

Exposure Rating

# Compute rate per exposure
rate = af.compute_rate_per_exposure(
    pure_premium=policies['pure_premium'],
    exposure=policies['exposure'],
    loadings={'profit': 0.05, 'expenses': 0.15}
)

# Create class plan
relativities = {
    'age_group': {'18-25': 1.5, '26-35': 1.0, '36-45': 0.8, '46+': 0.7},
    'vehicle_type': {'sedan': 1.0, 'suv': 1.2, 'sports': 1.8}
}

rated_policies = af.create_class_plan(
    data=policies,
    rating_factors=['age_group', 'vehicle_type'],
    base_rate=100.0,
    relativities=relativities,
    exposure_col='exposure'
)

# Apply trend factor
trended_losses = af.apply_trend_factor(
    historical_value=100000,
    trend_rate=0.03,
    years=4
)
print(f"Trended to current: ${trended_losses:,.0f}")

Portfolio Impact Analysis

# Premium impact from factor changes
factor_changes = {
    'age_group': {'18-25': 1.10, '26-35': 1.00, '36+': 0.95}
}

impact = af.compute_premium_impact(
    data=policies,
    base_premium_col='premium',
    factor_changes=factor_changes
)

print(f"Total impact: ${impact['premium_change'].sum():,.0f}")
print(f"Average change: {impact['premium_change_pct'].mean():.1f}%")

# Segment analysis
segment_impact = af.segment_impact_analysis(
    data=impact,
    premium_col='premium_current',
    segment_cols=['age_group', 'territory'],
    proposed_premium_col='premium_proposed'
)

# Price elasticity
elasticity = af.estimate_demand_elasticity(
    price_history=historical_data,
    price_col='average_premium',
    volume_col='policy_count'
)
print(f"Elasticity: {elasticity['elasticity']:.3f}")

# Elasticity curve
curve = af.compute_elasticity_curve(
    current_price=100,
    current_volume=10000,
    elasticity=-1.5,
    price_range=(80, 120)
)

Documentation

Full documentation available at: https://actuaflow.readthedocs.io

• API Reference
• User Guide
• Examples

Examples

Jupyter notebooks demonstrating common workflows:

• examples/motor_insurance.ipynb - Complete motor insurance pricing
• examples/property_pricing.ipynb - Property insurance GLMs
• examples/glm_basics.ipynb - GLM fundamentals
• examples/exposure_rating.ipynb - Exposure-based rating
• examples/portfolio_analysis.ipynb - Premium impact and elasticity

Package Structure

actuaflow/
├── glm/                  # GLM models and variants
│   ├── models.py         # BaseGLM, FrequencyGLM, SeverityGLM, TweedieGLM
│   ├── diagnostics.py    # VIF, lift curves, Gini, etc.
│   ├── gam.py            # GAM (smooth non-linear effects)
│   ├── monotone_glm.py   # MonotoneGLM (isotonic monotonicity constraints)
│   ├── quantile_regression.py # QuantileGLM (conditional quantile regression)
│   ├── qpp.py            # QPP pricing (FrequencyModel, QPPAdjuster, QPPPricer)
│   ├── bayesian_glm.py   # BayesianGLM (PyMC-based, optional `pymc` dependency)
│   ├── mixed_models.py   # HierarchicalGLM (mixed effects)
│   ├── zero_inflated.py  # ZeroInflatedPoisson, ZeroInflatedNegativeBinomial
│   └── hurdle.py         # HurdleGamma, HurdleLognormal
├── freqsev/              # Frequency-severity workflow
│   ├── frequency.py      # FrequencyModel
│   ├── severity.py       # SeverityModel
│   ├── aggregate.py      # AggregateModel, premium calculation
│   ├── copula.py         # FrequencySeverityCopula (dependence modeling)
│   └── large_loss.py     # LargeLossHandler
├── exposure/             # Exposure rating tools
│   ├── rating.py         # Class plans, rate per exposure
│   ├── trending.py       # Trend factors, on-leveling
│   └── exposure_curves.py # ExposureCurve
├── portfolio/            # Portfolio analysis
│   ├── impact.py         # Premium impact, mix-shift
│   ├── elasticity.py     # Price elasticity, optimization
│   └── lifetime_value.py # CustomerLifetimeValue
├── tariff/               # Tariff plans and rating engine
│   ├── factor_table.py   # FactorTable
│   ├── tariff_plan.py    # TariffPlan (base rate, tiers, eligibility, serialization)
│   └── rating_engine.py  # RatingEngine, build_tariff_from_glm
├── optimization/         # Rate optimization
│   └── rate_optimizer.py # ConstrainedRateOptimizer, dislocation analysis
├── compliance/           # Regulatory filing support
│   ├── rate_indication.py # RateIndicationSummary
│   ├── rate_filing.py    # RateFilingPackage, anti-discrimination checks
│   └── serff_export.py   # SERFFExporter (XML/CSV)
├── diagnostics/          # Actuarial diagnostics
│   ├── diagnostics.py    # ActuarialDiagnostics, loss_ratio_validation, etc.
│   └── actuarial.py      # actual_vs_expected, double_lift_chart, minimum_bias_check
├── credibility/          # Credibility weighting
│   ├── buhlmann.py       # BuhlmannCredibility
│   └── limited_fluctuation.py # LimitedFluctuationCredibility
├── data/                 # Data quality
│   └── quality.py        # DataQualityChecker, AutoImputer
├── registry/             # Model persistence
│   └── model_store.py    # ModelStore
├── explain/              # Model interpretability
│   └── change_explainer.py # ChangeExplainer
├── assumptions/          # Assumption management
│   ├── projection.py     # AssumptionProjector, InflationAdjuster
│   └── assumption_helper.py # AssumptionHelper
├── config/                # Configuration
│   └── settings.py       # settings
├── core/                 # Event bus and fitting strategy
│   ├── events.py         # ModelFittedEvent, ModelSavedEvent
│   └── strategy.py       # FittingStrategy
└── utils/                # Utilities
    ├── data.py           # Data loading, preparation
    ├── validation.py     # Input validation
    ├── cross_validation.py # TimeSeriesSplit, cross_val_score
    ├── context.py         # model_fitting_context
    └── retry.py           # retry_on_convergence

Key Concepts

Frequency-Severity Approach

The package follows the standard actuarial practice of modeling claim frequency and severity separately:

• Frequency Model: Models claim counts per policy (Poisson or Negative Binomial)
• Severity Model: Models claim amounts given a claim occurred (Gamma or Lognormal)
• Pure Premium: Frequency × Severity
• Loaded Premium: Pure Premium × Loading Factors

Offset in Frequency Models

Exposure is handled as an offset (not a predictor):

# Correct - offset parameter
freq_model.fit(data, formula='claim_count ~ age + vehicle', offset='exposure')

# Incorrect - don't include offset in formula
# freq_model.fit(data, formula='claim_count ~ age + vehicle + offset(log(exposure))')

Factor Relativities

For log-link models, relativities are exp(coefficient):

relativities = freq_model.get_relativities()
# age_group[18-25]: 1.50 means 50% higher frequency than base
# age_group[36-45]: 0.80 means 20% lower frequency than base

Best Practices

1. Always check dispersion for Poisson models - use Negative Binomial if overdispersed
2. Filter zero claims when fitting severity models
3. Use proper offsets for frequency models with varying exposure
4. Check VIF for multicollinearity before finalizing model
5. Validate on holdout data using lift curves and Gini
6. Document model changes using export functions
7. Monitor actual vs expected regularly

Requirements

• Python >= 3.9
• pandas >= 2.0.0
• polars >= 0.19.0 (for fast data loading)
• numpy >= 1.24.0
• scipy >= 1.10.0
• statsmodels >= 0.14.0
• scikit-learn >= 1.3.0

Contributing

ActuaFlow is a solo-maintained project. External code contributions are not accepted at this time. However, feedback, bug reports, and feature requests are always welcome!

To report issues or request features:

• Email: michael@watsondataandrisksolutions.com
• Subject: [BUG REPORT] for bugs or [FEATURE REQUEST] for ideas

See CONTRIBUTING.md for details.

License

Mozilla Public License v2.0 (MPL-2.0)

ActuaFlow is licensed under the Mozilla Public License v2.0, an OSI-approved open-source license that:

• ✅ Allows commercial and research use
• ✅ Permits use in proprietary applications
• ✅ Requires attribution and license inclusion

Special Licensing Available: Implementation services, consulting, GUI software, and alternative licensing terms. Contact the author for details.

See LICENSING.md for comprehensive licensing information.

Author: Michael Watson (michael@watsondataandrisksolutions.com)
Copyright: (c) 2026–present Michael Watson

For complete details, see:

• LICENSE - Full legal text
• LICENSING.md - Comprehensive usage guide
• CONTRIBUTING.md - Contributor guidelines

Licensing Options:

• Open Source: MPL-2.0 (free for research and commercial use)
• Commercial: Custom licensing available - contact michael@watsondataandrisksolutions.com

Citation

If you use ActuaFlow in academic research, please cite:

@software{actuaflow2025,
  title = {ActuaFlow: Modern Actuarial Pricing Library},
  author = {Watson, Michael},
  year = {2025},
  url = {https://github.com/WattyAI/actuaflow},
  license = {MPL-2.0}
}

Support

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Documentation: ReadTheDocs

Acknowledgments

ActuaFlow is built on:

• statsmodels for GLM implementation
• polars for fast data processing
• Industry best practices from CAS monographs and academic research

---

ActuaFlow - Modern tools for modern actuaries