How patrickresearch/forecastlens works
ForecastLens is a framework-agnostic diagnostic layer for existing time-series forecasts; it does not train models. The main workflow normalizes model output into a shared forecast representation, then evaluates accuracy overall or by detected regime, or translates forecasts into decision outcomes. Separate capabilities include leak-safe walk-forward splitting and synthetic price/demand data with ground truth. The web demo illustrates regime-aware evaluation and procurement timing. Edges follow README statements or source-index references; unsampled implementations are represented by their exact paths, but their internal calls are not inferred.
Main entry points: result.py ForecastResult · regime_aware.py Regime evaluator · bucket_accuracy.py Bucket accuracy · app.py Interactive demo
Components 21
Forecast Inputs
- Darts adapter
src/forecastlens/adapters/darts.py - NeuralForecast adapter
src/forecastlens/adapters/neuralforecast.py - Tabular adapter
src/forecastlens/adapters/raw.py - ForecastResult
src/forecastlens/core/result.py - Walk-forward splitter
src/forecastlens/backtesting/splitter.py
Evaluation
- Forecast metrics
src/forecastlens/metrics/pinball.py - Demand evaluator
src/forecastlens/evaluators/demand.py - Regime evaluator
src/forecastlens/evaluators/regime_aware.py - Regime detectors
src/forecastlens/regime/cusum.py - Regime clustering
src/forecastlens/regime/clustering.py - Volatility detector
src/forecastlens/regime/volatility.py
Decision Analysis
- Decision buckets
src/forecastlens/decision/buckets.py - Decision costs
src/forecastlens/decision/cost_matrix.py - Bucket accuracy
src/forecastlens/decision/bucket_accuracy.py - Value score
src/forecastlens/decision/value_score.py - Procurement timing
src/forecastlens/decision/procurement_timing.py
Data and Demo
- Synthetic price data
src/forecastlens/synthetic/price.py - Synthetic demand data
src/forecastlens/synthetic/demand.py - Generated series
src/forecastlens/synthetic/series.py - Interactive demo
web_demo/app.py
Other components
- Forecast user
Connections 23
- Forecast user → Darts adapter: provides forecast
- Forecast user → NeuralForecast adapter: provides forecast
- Forecast user → Tabular adapter: provides tabular output
- Darts adapter → ForecastResult: normalizes to
src/forecastlens/adapters/darts.py - NeuralForecast adapter → ForecastResult: normalizes to
src/forecastlens/adapters/neuralforecast.py - Tabular adapter → ForecastResult: normalizes to
src/forecastlens/adapters/raw.py - ForecastResult → Regime evaluator: evaluated by
src/forecastlens/evaluators/regime_aware.py - ForecastResult → Demand evaluator: evaluated by
src/forecastlens/evaluators/demand.py - ForecastResult → Procurement timing: informs timing
src/forecastlens/decision/procurement_timing.py - Forecast user → Walk-forward splitter: requests splits
- Forecast user → Forecast metrics: scores forecast
- Regime evaluator → Regime detectors: detects regimes
src/forecastlens/evaluators/regime_aware.py - Regime evaluator → Forecast metrics: computes scores
src/forecastlens/evaluators/regime_aware.py - Regime clustering → Regime detectors: clusters detections
src/forecastlens/regime/clustering.py - Volatility detector → Regime detectors: implements detector
src/forecastlens/regime/volatility.py - Bucket accuracy → Decision buckets: assigns buckets
src/forecastlens/decision/bucket_accuracy.py - Bucket accuracy → Decision costs: prices outcomes
src/forecastlens/decision/bucket_accuracy.py - Bucket accuracy → Value score: computes value
src/forecastlens/decision/bucket_accuracy.py - Synthetic price data → Generated series: returns series
src/forecastlens/synthetic/price.py - Synthetic demand data → Generated series: returns series
src/forecastlens/synthetic/demand.py - Interactive demo → Synthetic price data: loads presets
web_demo/app.py - Interactive demo → Regime evaluator: runs evaluation
web_demo/app.py - Interactive demo → Procurement timing: runs timing
web_demo/app.py