R

Default: current R release, renv for package versions, scripts run with Rscript, base stats for time series (ts, stl, decompose), ggplot2 for charts, testthat for tests, a rocker/r-ver image pinned to the R version for reproducible runs.

Setup

1. renv::init() in the project, commit renv.lock; collaborators run renv::restore().
2. Put analysis code in functions under R/ and keep top-level scripts short (Rscript run.R input.csv).
3. Pin the Docker base image to the same R version as renv.lock. Automated dependency bots do not read renv.lock, so let them track the base image tag instead and update packages deliberately.

Monthly time series

• Build a complete monthly index first; missing months must be explicit NA, not absent rows, or seasonal lags shift.
• ts(x, frequency = 12, start = c(year, month)), then stl(x, s.window = "periodic") for trend + seasonal + remainder (needs at least two full years and no NA — impute or trim first).
• A centred 2×12 moving average (stats::filter(x, c(1/24, rep(1/12, 11), 1/24), sides = 2)) shows trend without seasonality; a plain 12-term window is off-centre by half a month.
• Flag outliers on the remainder, not the raw series (e.g. |remainder| > 3 × MAD); a winter peak is not an outlier.
• Report each flag with the month, value, expected value and the rule that fired.

Data handling

• Read with explicit column types (readr::read_csv(col_types = ...) or colClasses); check problems().
• Dates with as.Date(x, format = "%Y-%m-%d"); avoid locale-dependent month names.
• Keep data tidy (one row per observation) before plotting or modelling.

Charts

• ggplot2 with explicit scales and units in axis titles; save with ggsave(width, height, dpi) so output is identical across machines.

Testing

• testthat for functions: complete-index creation, imputation, outlier rule on a synthetic series with a planted spike.
• Set seeds (set.seed) wherever randomness is used, and record them.

Verification loop

1. Rscript -e 'renv::status()' reports a consistent library.
2. Rscript -e 'testthat::test_dir("tests")'.
3. Run the script in the pinned container and compare outputs with the local run.

Pitfalls

• Factors from strings: older code and some readers convert strings to factors; check str(df).
• NA propagation: mean(x) is NA if any value is missing; decide between na.rm = TRUE and imputation explicitly.
• 1-based indexing and recycling: vectors of different lengths recycle silently, often with only a warning.
• sapply return types change with input; use vapply with a declared type.
• Floating comparisons: use all.equal or a tolerance, not ==.