# Models for mixed frequency time series Fit a structured time series to mixed frequncy data. 1.0 ```r bsts.mixed(target.series, predictors, which.coarse.interval, membership.fraction, contains.end, state.specification, regression.prior, niter, ping = niter / 10, seed = NULL, truth = NULL, ...) ``` ## Arguments - `target.series`: A vector object of class `zoo` indexed by calendar dates. The date associated with each element is the LAST DAY in the time interval measured by the corresponding value. The value is what Harvey (1989) calls a 'flow' variable. It is a number that can be viewed as an accumulation over the measured time interval. - `predictors`: A matrix of class `zoo` indexed by calendar dates. The date associated with each row is the LAST DAY in the time interval encompasing the measurement. The dates are expected to be at a finer scale than the dates in `target.series`. Any predictors should be at sufficient lags to be able to predict the rest of the cycle. - `which.coarse.interval`: A numeric vector of length `nrow(predictors)` giving the index of the coarse interval corresponding to the end of each fine interval. - `membership.fraction`: A numeric vector of length `nrow(predictors)` giving the fraction of activity attributed to the coarse interval corresponding to the beginning of each fine interval. This is always positive, and will be 1 except when a fine interval spans the boundary between two coarse intervals. - `contains.end`: A logical vector of length `nrow(predictors)` indicating whether each fine interval contains the end of a coarse interval. - `state.specification`: A state specification like that required by `bsts`. - `regression.prior`: A prior distribution created by `SpikeSlabPrior`. A default prior will be generated if none is specified. - `niter`: The desired number of MCMC iterations. - `ping`: An integer indicating the frequency with which progress reports get printed. E.g. setting `ping = 100` will print a status message with a time and iteration stamp every 100 iterations. If you don't want these messages set `ping < 0`. - `seed`: An integer to use as the random seed for the underlying C++ code. If `NULL` then the seed will be set using the clock. - `truth`: For debugging purposes only. A list containing one or more of the following elements. If any are present then corresponding values will be held fixed in the MCMC algorithm. * A matrix named `state` containing the state of the coarse model from a fake-data simulation. * A vector named `beta` of regression coefficients. * A scalar named `sigma.obs`. - ``...``: Extra arguments passed to SpikeSlabPrior ## Returns An object of class `bsts.mixed`, which is a list with the following elements. Many of these are arrays, in which case the first index of the array corresponds to the MCMC iteration number. - **coefficients**: A matrix containing the MCMC draws of the regression coefficients. Rows correspond to MCMC draws, and columns correspond to variables. - **sigma.obs**: The standard deviation of the weekly latent observations. - **state.contributions**: A three-dimensional array containing the MCMC draws of each state model's contributions to the state of the weekly model. The three dimensions are MCMC iteration, state model, and week number. - **weekly**: A matrix of MCMC draws of the weekly latent observations. Rows are MCMC iterations, and columns are weekly time points. - **cumulator**: A matrix of MCMC draws of the cumulator variable. The returned object also contains MCMC draws for the parameters of the state models supplied as part of `state.specification`, relevant information passed to the function call, and other supplemental information. ## References Harvey (1990), "Forecasting, structural time series, and the Kalman filter", Cambridge University Press. Durbin and Koopman (2001), "Time series analysis by state space methods", Oxford University Press. ## Author(s) Steven L. Scott steve.the.bayesian@gmail.com ## See Also `bsts`, `AddLocalLevel`, `AddLocalLinearTrend`, `AddSemilocalLinearTrend`, `SpikeSlabPrior`, `SdPrior`. ## Examples ```r ## Not run: data <- SimulateFakeMixedFrequencyData(nweeks = 104, xdim = 20) ## Setting an upper limit on the standard deviations can help keep the ## MCMC from flying off to infinity. sd.limit <- sd(data$coarse.target) state.specification <- AddLocalLinearTrend(list(), data$coarse.target, level.sigma.prior = SdPrior(1.0, 5, upper.limit = sd.limit), slope.sigma.prior = SdPrior(.5, 5, upper.limit = sd.limit)) weeks <- index(data$predictor) months <- index(data$coarse.target) which.month <- MatchWeekToMonth(weeks, months[1]) membership.fraction <- GetFractionOfDaysInInitialMonth(weeks) contains.end <- WeekEndsMonth(weeks) model <- bsts.mixed(target.series = data$coarse.target, predictors = data$predictors, membership.fraction = membership.fraction, contains.end = contains.end, which.coarse = which.month, state.specification = state.specification, niter = 500, expected.r2 = .999, prior.df = 1) plot(model, "state") plot(model, "components") ## End(Not run) ```

mixed.frequency function