Skip to contents

Changelog

Source: NEWS.md

bridgr 1.0.0

First stable release. Since 0.1.2 the model-construction entry point and the fitted-model class have been renamed, the aggregation library and the uncertainty machinery have been substantially extended, and two methods that did not honour their documented contract have been corrected. The public interface centred on mf_model(), forecast(), summary() and plot() is now considered stable, and future breaking changes will go through a deprecation cycle.

Breaking changes

  • Rename the main model-construction entry point to mf_model(), and rename the fitted-model class and its S3 methods from bridge to mf_model. bridge() remains as a deprecated compatibility wrapper that warns and forwards to mf_model(), so existing code keeps working.

  • summary() on an "mf_model" object now returns a "summary.mf_model" object instead of printing and returning the model unchanged, following the convention of summary.lm(). The printed report is unchanged and is now produced by the new print.summary.mf_model() method. The returned object exposes the summary quantities programmatically, including a standard coefficients matrix with Estimate, Std. Error, t value and Pr(>|t|) columns, so coef(summary(model)) works as it does for lm(). Standard errors respect the HAC, Delta-HAC or bootstrap covariance when the model was fitted with se = TRUE.

  • Objects returned by forecast() no longer inherit from the forecast package’s "forecast" class; they are now plain "mf_model_forecast" objects. The previous inheritance was not honoured – plot() and autoplot() failed on the result, and accuracy() returned misleading values – because target frequencies such as daily and weekly cannot be represented by stats::ts(), which those methods require. plot() and ggplot2::autoplot() methods are now provided directly for "mf_model_forecast" and work at every supported target frequency, and the new as.forecast() converts to a genuine "forecast" object for use with functions such as forecast::accuracy() whenever the target frequency has an exact ts representation (annual, semi-annual, quarterly, bi-monthly or monthly).

  • Remove the legendre parametric aggregation option.

New features

  • New accessor methods replace reaching into the fitted object’s internal structure: weights() returns aggregation weights (estimated parametric weights or user-supplied numeric weights), aggregation_parameters() returns the estimated parameters of parametric aggregation schemes, indicators() returns the indicator names, variable.names() returns the bridge-equation regressor names, and model.frame() returns the estimation data or the forecast regressor path. weights() and aggregation_parameters() accept an indicator name or position. The vignettes now use these accessors throughout.

  • variable.names() replaces model$xreg_names and model$regressor_names. variable.names(model, which = "xreg") returns the non-target-lag regressors, which are exactly the series a custom xreg must supply when forecasting a scenario, and so pairs with model.frame(model, which = "forecast").

  • weights() now also returns the fixed weight vectors implied by the deterministic aggregators, rather than NULL: "mean" gives 1/M, "last" gives a one in the final slot, and "sum" gives ones. The accessor therefore reports the weights actually applied for every aggregator except "unrestricted", which estimates one coefficient per within-period observation and so implies no weight vector.

  • Extend mf_model() beyond classic bridge aggregation:

    • add unrestricted mixed-frequency regressors via indic_aggregators = "unrestricted"
    • add parametric "beta" weighting alongside "expalmon"
    • add direct high-frequency alignment via indic_predict = "direct"
    • support fixed numeric aggregation weights supplied in a list()

  • Improve mixed-frequency input handling:

    • infer regular frequencies from second through year
    • allow custom frequency_conversions
    • standardize month-, quarter-, and year-end dates to period starts when needed for frequency recognition
    • keep the most recent observations in overfilled target periods with a summarized warning
    • fail when target periods contain too few high-frequency observations

  • Add joint parametric aggregation optimization controls through solver_options, including optimizer choice, multi-start runs, seeds, iteration limits, and user-supplied starting values.

  • Add uncertainty support:

    • se = TRUE for coefficient uncertainty and prediction intervals
    • HAC standard errors for linear bridge equations
    • Delta-HAC standard errors when parametric aggregation weights are estimated jointly
    • residual-resampling prediction intervals by default
    • optional full-system block bootstrap uncertainty via full_system_bootstrap = TRUE

  • Add scenario forecasting support in forecast.mf_model() through custom future xreg paths and standardized forecast objects with uncertainty metadata.

  • Add plotting methods and helpers:

  • Expand printed output and documentation:

    • standardize summary.mf_model() and forecast.mf_model() output
    • add vignettes on mixed-frequency modeling, ragged-edge nowcasting, and uncertainty / scenario analysis
    • refresh the README examples and package references

Performance

  • The full-system block bootstrap is substantially faster. Calendar shifts were applied one step at a time and recomputed for every observation, which made timezone normalisation inside lubridate::%m+% the dominant cost of a bootstrap resample. Shifts are now vectorised and computed once per distinct shift amount. On a quarterly target with a monthly indicator, a 50-draw full-system bootstrap runs about 3.6 times faster, with bit-identical coefficients and forecasts.

  • Month, quarter and year shifts of Date vectors no longer go through lubridate::%m+%, which routes through as.POSIXlt() and force_tz(). Profiling showed that timezone coercion alone accounted for roughly 44% of the remaining self time in a full-system bootstrap. These shifts now use direct integer calendar arithmetic, preserving the end-of-month rollback semantics of %m+% exactly; POSIXct inputs, missing values and fractional shifts still use %m+%. The isolated shift is about 10 times faster and a 50-draw full-system bootstrap about 1.3 times faster, with bit-identical coefficients, covariances, forecasts and intervals.

  • Use analytic gradients for expalmon optimization and improve the normalized beta polynomial gradient used in the optimizer.

Bug fixes

  • Fix mf_model() failing when the ‘xts’ package is not installed. Indicator forecasting with indic_predict = "auto.arima" (the default) or "ets" routed the series through tsbox::ts_xts(), which requires ‘xts’, so the default code path errored for users without it even though ‘xts’ was only a suggested dependency. The fitters are now given the indicator observations directly. Results are unchanged: the ‘xts’ index carried no tsp attribute, so both fitters already saw a frequency-1 series at every supported indicator frequency. ‘xts’ is no longer a dependency of any kind.

  • Fix ragged-edge completion for sub-monthly indicators at multi-step horizons (h > 1). Completion previously filled future target periods with a fixed count of high-frequency grid steps, but calendar periods can hold more observations than the regular ladder implies (a quarter has 13 weekly or up to 92 daily observations versus the 12 or 84 the ladder expects), so early future periods absorbed the surplus and later ones failed block validation. Completion is now period-aware: candidate grid times are assigned to their calendar periods and each future period receives exactly the observations it still needs.

  • Ignore indicator observations dated beyond the last forecast period during alignment. Such observations cannot enter any regressor and previously made block validation fail on a partially observed beyond-horizon period, for example when a weekly series extends past the target quarter of an h = 1 nowcast.

  • Make direct alignment (indic_predict = "direct") period-aware. Blocks of high-frequency observations were strided backward from the end of the sample and paired with target periods by position, so on calendar ladders (13-Saturday quarters on a 12-slot weekly ladder) historical blocks drifted out of their calendar periods – about one week per quarter, compounding over the sample – and the stride count could overrun the number of target periods and fail outright. Each target period that overlaps the observed sample is now anchored at the newest observation’s position within its own period (a MIDAS-with-leads alignment), which reproduces fixed strides exactly on regular ladders; periods beyond the observed sample keep the documented lead convention.

bridgr 0.1.2

CRAN release: 2026-02-18

  • Solve dependency issues with xts

bridgr 0.1.1

CRAN release: 2024-12-13

  • Initial CRAN submission:

    • Added gdp,baro, wea and fcurve datasets.

    • Added bridge(), forecast() and summary() functions.

    • Supports target variables on monthly, quarterly and yearly frequency, and indicator variables on daily, weekly, monthly, quarterly and yearly frequency.

    • Supports auto.arima, ets and other methods for indicator variable forecasting.

    • Supports aggregation of indicator variables to match the target’s frequency using custom weighting functions, exponential Almon polynomials and other methods.