PhilanthroPy: Code for a cause—predictive analytics for advancement teams.

🚀 View the Full Documentation Site

---

What is PhilanthroPy?

PhilanthroPy is a production-ready Python library that slots directly into sklearn.pipeline.Pipeline. It covers the full predictive workflow for nonprofit and academic medical center (AMC) fundraising — from raw CRM cleaning and wealth imputation to major-gift propensity scoring, lapse prediction, and planned-giving intent.

---

Installation

pip install philanthropy

From source (for development)

git clone https://github.com/PhilanthroPy-Project/PhilanthroPy.git
cd PhilanthroPy
pip install -e ".[dev]"

Or with Conda:

conda env create -f environment.yml && conda activate Philanthropy
pip install -e ".[dev]"

---

Quick Start

from philanthropy.datasets import generate_synthetic_donor_data
from philanthropy.models import DonorPropensityModel

df = generate_synthetic_donor_data(n_samples=500, random_state=42)
X = df[["total_gift_amount", "years_active", "event_attendance_count"]].to_numpy()
y = df["is_major_donor"].to_numpy()

model = DonorPropensityModel(n_estimators=200, random_state=0)
model.fit(X, y)
scores = model.predict_affinity_score(X)   # 0–100 affinity scale

Runnable scripts: examples/quickstart.py and examples/unischema_to_scores.py run end to end and are smoke-tested in CI.

Output of plot_affinity_distribution(): the 0–100 affinity scores cleanly separate major from non-major donors.

---

From UniSchema events to scores

PhilanthroPy is the modeling half of an ecosystem. UniSchema normalizes fragmented advancement webhooks (GiveCampus, Slate, NPSP, Cvent, …) into a single ConstituentEvent stream. philanthropy.ingest turns that stream into the donor-level feature table the estimators expect — no glue code between the two projects:

from philanthropy.ingest import read_constituent_events, constituent_events_to_features
from philanthropy.models import DonorPropensityModel

# 1. Load whatever UniSchema egressed — a .ndjson batch, a .json file, or a directory
events = read_constituent_events("data/egress/")

# 2. Aggregate to one row per donor — columns match the Quick Start above
features = constituent_events_to_features(events)

# 3. Score with a model trained on your labelled giving history (X_train, y_train).
#    Runnable end-to-end version: examples/unischema_to_scores.py
X = features[["total_gift_amount", "years_active", "event_attendance_count"]].to_numpy()
model = DonorPropensityModel(n_estimators=200, random_state=0).fit(X_train, y_train)
features["affinity_score"] = model.predict_affinity_score(X)

constituent_events_to_features is leakage-safe (recency anchored to an explicit reference_date or the batch's latest event, never a moving "now") and at-least-once-safe (deduplicates redelivered webhooks by eventId). Output, one row per constituent indexed by constituent_id:

Column	Built from
constituent_email, first_name, last_name	identity fields, carried through when present
total_gift_amount, gift_count, first_gift_date, last_gift_date	DONATION events
event_attendance_count	EVENT_REGISTRATION events
email_click_count	EMAIL_CLICK events
years_active, recency_days	first / last event vs. reference date
distinct_source_systems	channel breadth

---

Feature Overview — v0.4.0

🧹 Preprocessing

Transformer	Description
CRMCleaner	Standardise raw CRM exports — coerce types, strip whitespace, drop PII columns
WealthScreeningImputer	Leakage-safe wealth imputation (median / mean / zero) with optional missingness indicator columns; fill stats frozen at fit() time
WealthPercentileTransformer	Per-column wealth percentile rank (0–100); NaN-in → NaN-out
FiscalYearTransformer	Fiscal year & quarter features from gift dates; configurable fiscal year start month
RFMTransformer	Recency–Frequency–Monetary value feature engineering for donor segmentation
EncounterTransformer	Bridge hospital EHR encounter records with philanthropy CRM; computes days_since_last_discharge and encounter_frequency_score; snapshot at fit() prevents leakage
GratefulPatientFeaturizer ⭐ NEW	Clinical gravity score + AMC service-line capacity weights (cardiac 3.2×, oncology 2.9×, neuroscience 2.7×); outputs clinical_gravity_score, distinct_service_lines, distinct_physicians, total_drg_weight
DischargeToSolicitationWindowTransformer ⭐ NEW	Post-discharge solicitation window flags tracking recency; outputs in_solicitation_window (0/1), window_position_score (1.0 at midpoint, 0.0 at edges), and discharge_recency_tier (int 0-4); NaN → 0
PlannedGivingSignalTransformer ⭐ NEW	Bequest / legacy-gift intent vector: is_legacy_age, is_loyal_donor, inclination_score (−1 sentinel for absent data), composite_score [0–3]

🤖 Models

Model	Description
DonorPropensityModel	Random Forest classifier with predict_affinity_score() returning a 0–100 scale
MajorGiftClassifier	Calibrated HistGradientBoostingClassifier — NaN-native, with predict_affinity_score()
ShareOfWalletRegressor	Estimates total giving capacity and untapped-potential ratio
LapsePredictor	Random Forest classifier for donor lapse; see full docs below
MovesManagementClassifier	Multi-class portfolio stage predictor
PropensityScorer	Constant-probability baseline (P=0.5) — a floor to beat; use DonorPropensityModel for real scoring
PlannedGivingIntentScorer ⭐ NEW	Wraps GradientBoostingClassifier to predict bequest intent scores (0-100 scale)
FinancialForecastModel ⭐ NEW	Hybrid LSTM-ARIMA revenue/giving forecaster with predict_revenue_forecast(X, horizon); leakage-safe, dependency-free, sklearn-native

🔀 Model Selection

Splitter	Description
FiscalYearGroupedSplitter ⭐ NEW	Walk-forward cross-validator preventing fiscal year data leakage

📊 Metrics

Function	Description
donor_lifetime_value	LTV with configurable discount rate
retention_rate	Period-over-period donor retention
donor_acquisition_cost	CAC from campaign spend data

🗂 Datasets

Function	Description
generate_synthetic_donor_data	Reproducible synthetic prospect pool — n_samples, random_state

🔌 Ingest ⭐ NEW

Function	Description
constituent_events_to_features	Aggregate a UniSchema ConstituentEvent stream into a one-row-per-donor feature table; leakage-safe, deduplicates by eventId
read_constituent_events	Load UniSchema's JSON / NDJSON / directory egress into a list of event dicts

---

DischargeToSolicitationWindowTransformer

Post-discharge solicitation window featurization. Flags donors within the optimal window (days post-discharge) and emits proximity scores.

Parameter	Type	Default	Description
min_days_post_discharge	int	90	Start of solicitation window
max_days_post_discharge	int	365	End of solicitation window
days_since_discharge_col	str	"days_since_last_discharge"	Column name for days since discharge

Output columns: in_solicitation_window (0/1), window_position_score [0,1], discharge_recency_tier [0–4].

---

LapsePredictor

Production Random Forest classifier for donor lapse risk. Predicts whether a donor will lapse within a configurable window.

Parameter	Type	Default	Description
n_estimators	int	100	Number of trees
max_depth	int or None	None	Maximum tree depth
min_samples_leaf	int	1	Min samples per leaf
class_weight	dict/"balanced"/None	None	Class weighting
lapse_window_years	int	2	Prediction window in years
random_state	int or None	None	Reproducibility seed

Fitted attributes

Attribute	Description
estimator_	Trained RandomForestClassifier
classes_	Array of class labels
n_features_in_	Feature count from fit

Methods

Method	Returns	Description
.fit(X, y)	self	Train on features and binary lapse label
.predict(X)	ndarray (n,)	Binary predictions (1 = at-risk)
.predict_proba(X)	ndarray (n,2)	Class probabilities
.predict_lapse_score(X)	ndarray (n,)	P(lapse) × 100, range [0.0, 100.0]

from philanthropy.models import LapsePredictor

predictor = LapsePredictor(
    n_estimators=200,
    lapse_window_years=3,
    class_weight="balanced",
    random_state=42,
)
predictor.fit(X, y)
at_risk = predictor.predict(X)           # 1 = at risk of lapsing
scores  = predictor.predict_lapse_score(X)  # 0–100 lapse risk score

---

MajorGiftClassifier

Calibrated gradient-boosted classifier for major-gift propensity. Uses HistGradientBoostingClassifier wrapped in CalibratedClassifierCV for well-calibrated probabilities on skewed datasets. Handles NaN values natively — no upstream imputation needed.

Parameter	Type	Default	Description
learning_rate	float	0.1	Boosting step size
max_iter	int	100	Number of boosting trees
max_depth	int or None	None	Maximum tree depth
random_state	int or None	None	Reproducibility seed

Fitted attributes

Attribute	Description
estimator_	CalibratedClassifierCV wrapping HistGradientBoostingClassifier
classes_	Array of class labels
n_features_in_	Feature count from fit

Methods

Method	Returns	Description
.fit(X, y)	self	Train model
.predict(X)	ndarray (n,)	Binary predictions
.predict_proba(X)	ndarray (n,2)	Calibrated probabilities (rows sum to 1.0)
.predict_affinity_score(X)	ndarray (n,)	P(major gift) × 100, integer, range [0, 100]

Key distinction from DonorPropensityModel: DonorPropensityModel uses Random Forest with raw probabilities and float affinity scores. MajorGiftClassifier uses HistGradientBoosting + calibration, handles NaN natively, produces integer affinity scores, and provides better-calibrated probabilities on skewed datasets.

from philanthropy.models import MajorGiftClassifier
import numpy as np

model = MajorGiftClassifier(max_iter=200, random_state=42)
model.fit(X, y)

# Works with NaN values — no upstream imputation needed
X_with_nan = X.copy().astype(float)
X_with_nan[0, 1] = np.nan
scores = model.predict_affinity_score(X_with_nan)

---

PlannedGivingIntentScorer

Bequest/planned giving intent classifier. GradientBoostingClassifier + CalibratedClassifierCV for calibrated probabilities.

Parameter	Type	Default	Description
n_estimators	int	100	Boosting stages
random_state	int or None	None	Reproducibility seed

Methods: .fit(X, y), .predict(X), .predict_proba(X), .predict_intent_score(X) — P(intent) × 100, [0.0, 100.0].

---

FinancialForecastModel

Hybrid LSTM-ARIMA revenue/giving forecaster. Decomposes a giving series into a linear (ARIMA-surrogate) component fitted with LinearRegression and a nonlinear (LSTM-surrogate) residual component fitted with MLPRegressor — reproducing the additive hybrid y = linear(X) + nonlinear_residual(X) with no heavy dependencies (no TensorFlow / statsmodels). Multi-step forecasts roll a frozen autoregressive model forward. All fitted statistics (fill values, sub-models, AR coefficients) are frozen at fit() for leakage safety.

Parameter	Type	Default	Description
ar_order	int	3	Order of the autoregressive roll-forward
hidden_layer_sizes	tuple	(64,)	Residual network architecture (LSTM stand-in)
max_iter	int	300	Max optimisation iterations for the network
alpha	float	1e-4	L2 regularisation of the residual network
random_state	int or None	None	Reproducibility seed

Methods: .fit(X, y), .predict(X) → ndarray (n,), .predict_revenue_forecast(X, horizon) → ndarray (horizon,).

from philanthropy.models import FinancialForecastModel

model = FinancialForecastModel(ar_order=3, random_state=42)
model.fit(X, y)                                   # y = period giving revenue
forecast = model.predict_revenue_forecast(X, horizon=4)   # next 4 periods

---

API quick reference

Component	Module	Type
DonorPropensityModel	philanthropy.models	Classifier
MajorGiftClassifier	philanthropy.models	Classifier
LapsePredictor	philanthropy.models	Classifier
PlannedGivingIntentScorer	philanthropy.models	Classifier
ShareOfWalletRegressor	philanthropy.models	Regressor
FinancialForecastModel	philanthropy.models	Regressor
DischargeToSolicitationWindowTransformer	philanthropy.preprocessing	Transformer
RFMTransformer	philanthropy.preprocessing	Transformer

---

sklearn Pipeline Example

from sklearn.pipeline import Pipeline
from philanthropy.preprocessing import (
    FiscalYearTransformer, WealthScreeningImputer, DischargeToSolicitationWindowTransformer
)
from philanthropy.models import DonorPropensityModel

pipe = Pipeline([
    ("fy",      FiscalYearTransformer(date_col="gift_date")),
    ("wealth",  WealthScreeningImputer(wealth_cols=["estimated_net_worth"])),
    ("window",  DischargeToSolicitationWindowTransformer()),
    ("model",   DonorPropensityModel(n_estimators=200, random_state=0)),
])
pipe.fit(X_train, y_train)
scores = pipe.predict_proba(X_test)[:, 1]

---

Package Overview

philanthropy/
├── ingest/
│   └── _constituent_events.py   # UniSchema ConstituentEvent → donor features
├── datasets/
│   └── _generator.py
├── preprocessing/
│   ├── transformers.py
│   ├── _wealth.py
│   ├── _wealth_percentile.py
│   ├── _encounters.py
│   ├── _encounter_recency.py
│   ├── _rfm.py
│   ├── _discharge_window.py
│   ├── _solicitation_window.py   # alias → _discharge_window
│   ├── _grateful_patient.py
│   ├── _planned_giving.py
│   └── _share_of_wallet.py
├── models/
│   ├── propensity.py
│   ├── _propensity.py
│   ├── _wallet.py
│   ├── _lapse.py
│   ├── _planned_giving.py
│   ├── _moves.py
│   └── _forecast.py
├── metrics/
│   ├── scoring.py
│   └── _financial.py
├── model_selection/
│   └── __init__.py
├── experimental/
│   └── _lapse.py
├── visualisation/
│   └── _plots.py
└── utils/
    ├── testing.py
    └── _validation.py

---

Medical Philanthropy Pipeline

from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from philanthropy.preprocessing import EncounterTransformer, CRMCleaner

# 1. EHR data extraction
encounter_features = EncounterTransformer(
    encounter_df=encounter_df,
    encounter_date_col="discharge_date",
    donor_id_col="mrn"
)

# 2. Use ColumnTransformer to merge EHR features with numeric CRM data
# This prevents the silent dropping of original CRM columns like estimated_net_worth
preprocessor = ColumnTransformer(
    transformers=[
        ("encounters", encounter_features, ["mrn", "gift_date"]),
        ("crm_nums", CRMCleaner(), ["estimated_net_worth", "real_estate_value"]),
    ],
    remainder="passthrough"
)

# 3. Fit pipeline on original dataset (gift_df)
gift_features = preprocessor.fit_transform(gift_df)

---

Testing

1160 tests across 26 files — check_estimator compliance for every public estimator, property-based (Hypothesis) suites, and a dedicated temporal-leakage suite.

# Full suite
pytest tests/ -q

# sklearn check_estimator compliance
pytest tests/test_sklearn_compliance.py -q

# Property-based (Hypothesis)
pytest tests/test_properties.py tests/test_preprocessing_properties.py -q

# Leakage prevention
pytest tests/test_leakage.py -v

---

Roadmap

✅ Completed

• CRMCleaner — leakage-safe CRM standardisation
• WealthScreeningImputer — median/mean/zero imputation
• EncounterTransformer — clinical discharge → feature engineering
• RFMTransformer — Recency, Frequency, Monetary features
• DischargeToSolicitationWindowTransformer — solicitation window features
• ShareOfWalletRegressor — capacity regression + predict_capacity_ratio()
• MajorGiftClassifier — gradient-boosted with calibrated probabilities
• LapsePredictor — production RF with predict_lapse_score()
• PlannedGivingIntentScorer — planned giving intent classifier
• donor_lifetime_value() — NPV LTV with discount rate
• philanthropy.visualisation — affinity score plots
• Property-based Hypothesis testing for FiscalYearTransformer
• Temporal leakage prevention test suite
• GitHub Actions CI (Python 3.10 + 3.11 matrix)
• Coverage gate (≥ 85%)
• Makefile (make ci)
• Branch protection + PR workflow
• philanthropy.ingest — UniSchema ConstituentEvent → donor feature bridge
• MkDocs Material documentation site (GitHub Pages)
• PyPI release (pip install philanthropy)
• philanthropy.visualisation.plot_retention_waterfall()

🔜 Next

• philanthropy.visualisation.plot_capacity_heatmap()
• EnsemblePropensityModel (stacked LapsePredictor + DonorPropensityModel)

---

Design Principles

• Leakage-safe by design — fill statistics, encounter summaries, and encounter snapshots are all frozen at fit() time; transform() is fully idempotent
• sklearn-native — all estimators pass check_estimator; support set_output(transform="pandas"), clone(), get_params() / set_params()
• NaN-transparent — wealth and clinical transformers declare allow_nan = True; no silent data loss
• PII-aware — EncounterTransformer auto-drops PII-like columns before returning features

---

Contributing

See CONTRIBUTING.md for the full local test gate, the new-test-file workflow, and pre-push hook setup. In short: run make ci before every push, and never use git push --no-verify.

---

License

MIT License — see LICENSE for details.