ins-pricing 0.9.81

Reusable modelling, pricing, governance, and reporting utilities.

ins_pricing

An enterprise-grade Python library for insurance ML model training, pricing, model governance, and reporting.

Core dependencies are numpy and pandas only. Heavy ML dependencies (torch, optuna, xgboost, shap, etc.) are optional extras loaded lazily, so you can use the pricing, governance, and reporting modules without installing any ML packages.

Installation

pip install ins_pricing                    # core (numpy + pandas)
pip install ins_pricing[bayesopt]          # + torch, optuna, xgboost, glum, sklearn
pip install ins_pricing[explain]           # + torch, shap, sklearn
pip install ins_pricing[plotting]          # + matplotlib, sklearn
pip install ins_pricing[geo]              # + contextily, matplotlib
pip install ins_pricing[gnn]             # + torch, torch-geometric, pynndescent
pip install ins_pricing[frontend]         # + nicegui, openpyxl
pip install -e ".[dev,plotting,frontend]" # local development
pip install ins_pricing[all]             # everything

GPU note: Install the correct PyTorch build for your platform/GPU before installing extras. Torch Geometric requires platform-specific wheels. Multi-GPU training uses DDP (NCCL on Linux) or DataParallel; Windows falls back to Gloo.

Modules

Module	What it does
modelling	BayesOpt-driven training for GLM, XGBoost, ResNet, FT-Transformer, GNN. Includes evaluation (calibration, bootstrap CI), explainability (permutation, SHAP, integrated gradients), and plotting (lift, oneway, geo heatmaps).
pricing	Factor tables, numeric binning, exposure calculation, premium rating, calibration, data quality checks, PSI monitoring.
production	Inference registry (ModelSpec, PredictorRegistry), batch scoring, preprocessing pipelines, drift detection, production metrics.
governance	JSON-backed model registry, approval workflows, audit logging, release/rollback management.
reporting	Markdown report builder and daily scheduler.
frontend	NiceGUI web UI for config-driven training, explanation, plotting, prediction, and FT two-step workflows.
cli	Entry points: BayesOpt_entry.py/Pricing_Run.py (training), BayesOpt_incremental.py/Incremental_Run.py (incremental retraining), Explain_entry.py/Explain_Run.py (explanation), watchdog_run.py (auto-restart).
utils	Validation, loss resolution, device management, metrics, profiling, logging, safe pickle, path resolution.

Quick Start

Train models

from ins_pricing.modelling.bayesopt import BayesOptConfig
from ins_pricing.modelling import BayesOptModel

config = BayesOptConfig.from_file("config.json")
model = BayesOptModel(train_data, test_data, config=config)
model.optimize_model("xgb", max_evals=100)
model.optimize_model("resn", max_evals=50)
model.optimize_model("ft", max_evals=50)

Build pricing factors

from ins_pricing.pricing import compute_base_rate, build_factor_table, rate_premium

base_rate = compute_base_rate(df, loss_col="claim_amt", exposure_col="exposure")
age_factors = build_factor_table(df, factor_col="age_band", loss_col="claim_amt",
                                  exposure_col="exposure", base_rate=base_rate)
premium = rate_premium(df, exposure_col="exposure", base_rate=base_rate,
                        factor_tables={"age_band": age_factors})

Score in production

from ins_pricing.production import load_predictor_from_config, batch_score

predictor = load_predictor_from_config("config.json", "xgb", device="cuda")
scored = batch_score(predictor.predict, df, batch_size=10000)

Govern and release

from ins_pricing.governance import ModelRegistry, ReleaseManager

registry = ModelRegistry("registry/models.json")
registry.register("pricing_xgb", "v2", metrics={"rmse": 0.11})

release = ReleaseManager("registry/deployments", registry=registry)
release.deploy("prod", "pricing_xgb", "v2", actor="ops")
release.rollback("prod", actor="ops")   # revert to previous version

Generate reports

from ins_pricing.reporting import ReportPayload, write_report, schedule_daily

payload = ReportPayload(model_name="pricing_xgb", model_version="v2",
                         metrics={"rmse": 0.11, "loss_ratio": 0.63})
write_report(payload, "reports/monthly.md")
schedule_daily(lambda: write_report(payload, "reports/daily.md"), run_time="02:00")

CLI

# Training (single GPU)
python ins_pricing/cli/BayesOpt_entry.py --config-json config.json

# Training (multi-GPU DDP)
torchrun --nproc_per_node=2 ins_pricing/cli/BayesOpt_entry.py --config-json config.json

# Explanation
python ins_pricing/cli/Explain_entry.py --config-json config_explain.json

# NiceGUI frontend
python -m ins_pricing.frontend.app

All workflows are config-driven. The runner.mode field in the JSON config determines the task:

runner.mode	Task
entry	Model training
explain	Permutation importance, SHAP, integrated gradients
incremental	Incremental batch training
watchdog	Auto-restart monitoring

Project Structure

ins_pricing/
  modelling/
    bayesopt/           Training core: config, trainers, models, runtime
    explain/            Permutation importance, SHAP, integrated gradients
    plotting/           Lift curves, oneway, diagnostics, geo heatmaps
    evaluation.py       Calibration, threshold selection, bootstrap CI
  pricing/              Factor tables, exposure, calibration, data quality
  production/           Inference registry, scoring, preprocessing, monitoring
  governance/           Model registry, approval, audit, release management
  reporting/            Report builder and daily scheduler
  frontend/             NiceGUI web UI
  cli/                  CLI entry points and shared utilities
  utils/                Validation, losses, device, metrics, profiling, IO
tests/                  Tests mirroring package module structure
model_processing/       Demo notebooks and JSON config templates (not packaged)

Training Output Layout

Training writes to output_dir/ with three subdirectories:

output_dir/
  plot/       Diagnostic plots (oneway, lift, loss curves, geo)
  Results/    Metrics JSON, best params, evaluation snapshots
  model/      Saved model artifacts (XGB, ResNet, FT, GNN checkpoints)

Loss and Distribution

Set distribution in your config to control the loss function. It takes precedence over loss_name.

distribution	loss	XGBoost objective
tweedie	tweedie	reg:tweedie
poisson	poisson	count:poisson
gamma	gamma	reg:gamma
gaussian / normal / mse	mse	reg:squarederror
laplace / mae	mae	reg:absoluteerror
bernoulli / binomial / binary	logloss	binary:logistic

See ins_pricing/modelling/bayesopt/README.md for full details.

Development

See docs/development_setup.md for GitHub multi-device setup, shared Codex/Claude Code configuration, and the source-vs-artifact policy.

# Install in dev mode
pip install -e ".[dev,plotting,frontend]"

# Run tests
python -m pytest tests/ -v

# Build and publish
make build && make check && make upload

PyPI Upload

# Linux / macOS
export TWINE_PASSWORD='your_pypi_token'
make build && make upload

# Windows
set TWINE_PASSWORD=your_pypi_token
python -m build
upload_to_pypi.bat