[viniciusoike]

Adds a 12-month demand forecast to the Lines tab. The forecast appears when a single line is selected and the new "Mostrar projeção (12 meses)" checkbox is on.

Model

• Equal-weight ensemble of STL + ETS, ETS, and ARIMA with calendar regressors, fitted on log demand since Jan 2022, so the 2020–21 collapse and recovery stay out of the fit.
• Calendar regressors are the business days and non-holiday Saturdays per month, built from metrosp::calendar_spo.
• In a rolling-origin backtest, the ensemble's MAPE was 3.5%, against 4.1% for the best single model (ARIMA with calendar). Blanking the pandemic months and fitting from 2016 did worse (7.3%).

Intervals

• The 90% intervals are empirical. tools/forecast-backtest.R pools the ensemble's log errors by horizon across lines, after standardizing each by its series' residual SD, and writes the quantiles to data/forecast-quantiles.csv. The app reads that file at startup.
• The quantiles are forced to widen (or stay flat) with the horizon.
• Coverage is 92–97% for Lines 1–4, 86% for Line 15 and 81% for Line 5. Line 5 is biased upward.

App changes

• A fifth KPI card shows the projected 12-month growth with its interval.
• The Lines CSV includes fc_mean, fc_lower and fc_upper rows when the forecast is on.
• forecast and its dependencies are added to renv.lock.
• tools/check-manifest.R now also covers data/.

Dependencies

• renv.lock was re-snapshotted with r-universe listed as a repository, so metrosp and trendseries (now 1.7.0) are pinned to r-universe.
• manifest.json was regenerated. It now records both packages from r-universe, plus forecast and data/forecast-quantiles.csv.

🤖 Generated with Claude Code