Cumulative event returns explicitly reject returns below -1 in DuckDB, including DuckDB 2.0 with ieee_floating_point_ops enabled (#34).
farr 1.1.0
Expanded runnable examples and return-value documentation, documented the prediction method for rusboost() objects, and clarified which examples require internet access or database credentials.
get_size_rets_monthly() now downloads over HTTPS to a temporary file, removes it even on failure, and restores VROOM_CONNECTION_SIZE after parsing instead of changing the caller’s files or environment (#33).
rusboost() now retains a perfect learner as the sole predictive learner with weight 1 and reports an error if no usable learner is fitted (#32). Random undersampling handles singleton classes safely and validates class counts and majority sample sizes.
Monthly cumulative-return windows now include the full ending month for nonzero offsets, including transitions from shorter to longer months.
Daily and monthly cumulative event returns now handle complete losses (ret = -1) in stocks and benchmarks (#31). Missing-value handling is unchanged, and returns below -1 continue to raise an error.
get_event_rets() now retains non-trading-day announcements (#30), measuring relative trading days from the next trading day while keeping the original announcement dates.
Corrected get_got_data() to use rho_E for firm-specific error persistence (#29). Simulations with unequal rho_X and rho_E now follow the requested error process; equal-parameter results are unchanged.
Refreshed aaer_dates through October 2026, with 3,386 rows. The data steps apply source-backed date and coverage corrections, preserve historical entries absent from the current SEC index, and remove exact duplicates (#27).
Corrected five public-float dates in iliev_2010 using the original 10-Ks, including four fractional years. The original CSV is preserved, with an explicit correction table and filing references in data-raw (#28).
Corrected tree weights in predict.rusboost(type = "prob") (#26), including predictions for a single observation or weak learner. Probability predictions require finite, nonnegative weights with a positive finite total.
Corrected the by_tag_year documentation: year_total counts questions across all tags in a year, rather than questions with a specific tag.
get_ff_ind() now caches downloaded Fama-French industry zip files by default and supports refresh_cache = TRUE.
Added a pkgdown site with Bootstrap 5 styling, a dark-mode toggle, and a Quarto article scaffold.