tsbench

Cleaned benchmark snapshot for microbiome time-series forecasting.

The repo is organized around three layers:

• model backbones: GLV, LSTM/RNN, ODE models, MCP models, MicroFlow
• constraint handling: nonnegativity, simplex, grouped simplex, support caps
• validity auditing: raw, projected, and post-sanitize checks tracked separately

Recommended rerun order

pytest -q tests/test_refactor_smoke.py
SAVE_EXAMPLES=1 N_SPLITS=1 bash bash_scripts/run_compositional_refresh.sh
PLOT_EXAMPLES=1 bash bash_scripts/run_visualize.sh

Paper-facing core suite

pytest -q tests/test_refactor_smoke.py
SAVE_EXAMPLES=1 N_SPLITS=3 bash bash_scripts/run_paper_core_suite.sh
PLOT_EXAMPLES=1 bash bash_scripts/run_visualize.sh

See bash_scripts/README.md for the full script menu.