How will3watson3/pitch-intent-metric works
Pitch Intent Metric is a research estimator for plausible pitch finish locations from reviewed catcher-glove targets and earlier comparable pitches. The workflow validates and joins audit labels with pitch-level Statcast records, then estimates provisional finishes from historical target-to-finish offsets, abstaining when support is inadequate. The pilot runner also compares model variants and evaluates endpoint errors over held-out games. The README documents baselines and temporal evaluation; the sampled runner excerpt is partial, so only clearly supported connections are shown.
Main entry points: run_intent_pilot.py Pilot runner · build_webb_sweeper_intent_pilot.py Validate and join
Components 13
Data preparation
- Validate and join
scripts/build_webb_sweeper_intent_pilot.py - Model-ready pitches
Intent inference
- Pilot runner
scripts/run_intent_pilot.py - Comparable inference
scripts/run_intent_pilot.py - Provisional estimates
Pilot evaluation
- Variant predictions
- Baseline comparisons
scripts/run_intent_pilot.py - Temporal evaluation
scripts/run_intent_pilot.py - Pilot reports
Other components
- Researcher
- Reviewed labels
- Statcast records
- Pitch history
Connections 14
- Researcher → Validate and join: selects group
- Reviewed labels → Validate and join: reads labels
- Statcast records → Validate and join: reads records
- Validate and join → Model-ready pitches: writes accepted pitches
- Researcher → Pilot runner: runs pilot
- Pilot runner → Model-ready pitches: loads audited pitches
- Pilot runner → Pitch history: loads history
- Pilot runner → Comparable inference: calls inference
- Comparable inference → Provisional estimates: estimates finishes
- Pilot runner → Variant predictions: writes predictions
- Pilot runner → Baseline comparisons: compares methods
- Pilot runner → Temporal evaluation: evaluates held-out games
- Baseline comparisons → Pilot reports: reports metrics
- Temporal evaluation → Pilot reports: reports results