feat(modeling): add validated LR, SVGP and ST-SVGP pre-test pipeline

[Omer-alt]

Summary

This PR consolidates the probabilistic modeling phase for urban expansion forecasting in Yaoundé.

It adds and documents the retained Logistic Regression, SVGP and ST-SVGP workflows, together with the controlled feature-engineering experiments, mathematical validation gates, chronological rolling evaluation, calibration diagnostics and retained experimental alternatives.

The main objective of this branch is to provide a reproducible pre-test modeling pipeline .

---

What is included

Logistic Regression

• Retained chronological Logistic Regression baseline.
• Uses the same pre-test rolling validation protocol as the Gaussian Process models:
  • train 2000 → validate 2005
  • train 2000, 2005 → validate 2010
  • train 2000, 2005, 2010 → validate 2015
• The feature specification was updated after feature-engineering experiments.
• The interaction built_fraction_x_recent_growth_t was retained because it also improved the Logistic Regression comparator.

SVGP

The retained SVGP remains the Adam-based Bernoulli-probit sparse variational GP with:

• separable Matérn-3/2 spatial × temporal kernel;
• 64 fixed spatial inducing locations;
• spatial coordinates and time represented in the GP covariance rather than the parametric mean;
• chronological rolling validation.

Feature-engineering experiments

Three controlled feature sets were evaluated:

1. base
2. log_distance
3. log_distance_growth

The retained Strong SVGP is:

log_distance_growth

with the additional retained features:

log_distance_to_built_m_t
built_fraction_x_recent_growth_t

The current conclusion is therefore nuanced:

• Logistic Regression still leads on aggregate proper scoring rules.
• SVGP improves after feature engineering.
• ST-SVGP does not yet dominate overall.

---

Main files added / updated

Configuration

configs/modeling/st_svgp.yaml
configs/modeling/st_svgp/
configs/modeling/svgp_experiment.yaml
configs/modeling/logistic_regression_experiment.yaml

ST-SVGP implementation

src/models/train_st_svgp.py
src/models/st_svgp/state_space.py
src/models/st_svgp/filtering.py
src/models/st_svgp/cvi.py
src/models/st_svgp/spatial.py
src/models/st_svgp/block_inference.py
src/models/st_svgp/model.py

Mathematical validation

src/models/st_svgp/validate_temporal_kernel.py
src/models/st_svgp/validate_filter_smoother.py
src/models/st_svgp/validate_cvi_natgrad.py

Evaluation

src/models/evaluation/evaluate_probabilities.py
src/models/evaluation/calibration.py
src/models/evaluation/prediction_sets.py

Tests

tests/test_st_svgp.py
tests/test_st_svgp_filtering.py
tests/test_st_svgp_cvi.py
tests/test_st_svgp_model.py

Documentation

docs/modeling_st_svgp.md
docs/documented_file_structure.md
README.md
Makefile

---

Validation commands

The main checks can be reproduced with:

make st-svgp-tests
make st-svgp-pretraining-validation
make st-svgp-preflight
make st-svgp-rolling
make st-svgp-compare-oof

The promoted ST-SVGP candidate currently passes the retained validation gates and rolling workflow.

---

Out of scope

This PR intentionally does not include:

• final 2020 → 2025 evaluation;
• 2030 / 2035 production forecasts;
• row-level OOF Parquet prediction files;
• local archive directories;