Sprint forecast (exact-state, simulation-first)

This repository contains a reference implementation of the amended model:

• Supply / capacity: daily capacity is a product of planned hours, Bernoulli availability, an enumerated blocked/disruption state, and an unbounded lognormal productivity multiplier.
• Demand / workload: remaining work is modelled from three-point estimates with global + per-engineer bias correction on the log scale, then mapped to unbounded lognormal item-effort distributions.
• AI attribution: AI usage can be introduced as a covariate in both the demand-side (bias model / workload) and the supply-side (focus multipliers). The implementation computes counterfactual success probabilities with AI turned off.
• Bayesian pooling: conjugate pooling utilities (Beta--Binomial, Dirichlet--Multinomial) and linear-Gaussian shrinkage estimators for bias and focus parameters.

The core API is SprintSuccessForecaster.

Quickstart

Install:

python -m pip install sprint-forecast

Run the demo:

python -m sprint_forecast.demo

Zenhub sprint forecast

run_forecast.py supports pulling per-issue three-point estimates for the current open Zenhub sprint:

• Set workload_source.type to zenhub_sprint_issues (see config.example.json)
• Provide ZENHUB_API_TOKEN
• Export .scarf/pert-hours.json from the PERT Hours Tracker (or disable workload_source.pert.require_*)

Sprint process

See docs/sprint-process.md for the recommended workflow (assignments, PERT in effective hours, actual tracking, required fields, and AI assumptions).

Structure

• sprint_forecast/ – library code
• examples/ – runnable example
• tests/ – unit tests (python -m unittest)

Notes

• The implementation targets Python 3.11+.
• Randomness is handled via NumPy's Generator.