Multi-Model Time Series Forecasting Toolkit and Local App

Fits, compares and visualizes univariate time series forecasts through an extensible model registry built largely on the 'forecast' package: AutoRegressive Integrated Moving Average (ARIMA) and Seasonal ARIMA (SARIMA) models (Hyndman and Khandakar (2008) ), Error, Trend, Seasonal (ETS) exponential smoothing (Hyndman et al. (2002) ), Trigonometric seasonality, Box-Cox transformation, AutoRegressive Moving Average (ARMA) errors, Trend and Seasonal components (TBATS) models (De Livera, Hyndman and Snyder (2011) ), Neural Network AutoRegression (NNETAR) and Holt-Winters exponential smoothing (Winters (1960) ), plus optional 'Prophet' models via the 'prophet' package (Taylor and Letham (2018) ) and Long Short-Term Memory (LSTM) networks via the 'torch' package (Hochreiter and Schmidhuber (1997) ). Also provides grouped series with bottom-up and top-down reconciliation, rolling-origin cross-validation, ensembles and residual diagnostics, following Hyndman and Athanasopoulos (2021) < https://otexts.com/fpp3/>, and a bundled local 'shiny' application exposing every model.

forecastsuite

A local, full-featured companion to the hosted Forecast Dashboard Shiny app. Where the hosted app deliberately sticks to Prophet + ARIMA to fit a ~1GB-RAM free hosting tier, this package has no such constraint and adds:

A pluggable model registry (register_model()/get_model()/list_models()) — every model implements a fit/forecast/to_tibble contract, so adding a new model later is one file + one register_model() call.
Built-in models: Prophet (optional, requires prophet), ARIMA, SARIMA (both with auto and manual p,d,q/P,D,Q order entry), ETS, TBATS, NNETAR, Holt-Winters, and LSTM (optional, requires torch).
A bundled local Shiny app (run_app()) with all models exposed, Plotly always on, and a dataset-driven model recommendation step.
Import from any of four sources: a file upload (CSV, TSV or Excel -- multi-sheet workbooks let you pick the worksheet), a data frame already in your R global environment (or, esquisse-style, a built-in dataset from any installed package -- e.g. datasets, ggplot2 -- picked by package then dataset name, no library() or data() call needed first), a link-shared Google Sheet, a CSV URL, or pasted delimited text.
Dates split across columns (Year + Quarter, Year + Month, optionally + Day) are assembled automatically, and aggregation is set to the finest part supplied.
An N-way model comparison: overlay several models' forecasts on one Plotly chart alongside the side-by-side metrics table.
Population normalization: divide by a population table to forecast incidence (e.g. cases per 100,000) rather than absolute counts.
A full holiday system: Sundays, a fixed-date catalog, movable holidays from a file, manual entries, relabel/remove, per-holiday windows, and a consistency check against your data.
Multi-window evaluation (Train / Last 6 Months / Last 2 Years / held-out Test) and suggested starting hyperparameters (including ARIMA/SARIMA's p, d, q, P, D, Q, all data-derived) alongside the model recommendation.
Individual Observations as an alternative to Aggregated data: import one row per event (e.g. one case, one visit) and it's counted into periods for you, no value column needed. Population normalization works with either format.
A Model Guide tab explaining every model, the parameters, the metrics, and how the recommendation heuristic works.
A "Show Code" panel that prints runnable R reproducing whatever you just fit or compared, syntax-highlighted, downloadable as a .R file, or copyable to the clipboard in one click.
An Analysis panel (Model tab): seasonal decomposition (trend/seasonal/remainder), anomaly detection on the raw series (IQR or z-score, downloadable as CSV), residual diagnostics after a fit (Ljung-Box, Shapiro-Wilk, residual ACF, and prediction interval coverage -- is the error actually unpredictable noise, or is there structure left over, and are the uncertainty bounds honest?), bias/drift tracking (is the model's error systematically one-sided, and is it growing across the test window?), and, with a grouping column active, a cross-group correlation heatmap/table. Decomposition, anomaly detection, and bias/drift need only a finalized dataset or a fit, no extra setup.
Forecast ensembling: combine two or more compared models' forecasts (simple average or inverse-error-weighted) into one more entry on the same comparison plot/table -- "Build Ensemble from Selected" on the Model tab.
A multi-model backtest leaderboard: the same rolling-origin folds as Cross-Validation, but across every model selected for Compare Models at once, ranked by mean test MASE.
Plot appearance controls (trend/uncertainty/holiday/changepoint toggles, colour pickers for every line) and one-click downloads: the processed dataset and the holiday list as CSV, and the forecast plot -- single-model or the multi-model comparison -- as PNG.
A bundled 5-year demo dataset (simulated daily clinic visits, with missing and partially-open Sundays) so you can explore the whole app with no data of your own -- pick it from the same icon row as the other import sources.
An optional grouping column (District, Sex, Product, ...): split your dataset into one time series per distinct value instead of only ever seeing the aggregate, fit one model across every included group, browse each group's own plot/metrics, and compare them on one overlay plot (actuals + forecasts, one color per group). Population normalization, the recommendation table, and the holiday consistency check are all group-aware when it's on. Messy raw values ("female"/"Female"/ "FEMALE") are auto-merged by casing/spacing, plus a manual relabel table for anything else (abbreviations, typos) -- the same select-and-relabel pattern used for holidays.
Bottom-up hierarchical reconciliation: once you've fit two or more groups, view a "Reconciled (bottom-up)" forecast -- the sum of the already-fit group forecasts, always coherent by construction, labeled as partial if not every configured group was fit this run.
Rolling-origin cross-validation: refit the currently-viewed series at several walk-forward cutoffs instead of relying on one train/test split, for a more robust read on accuracy (Model tab, "Run Cross-Validation").
Save your dataset(s), holidays, and every setting as one project file, and reload it in a fresh session with no re-import needed -- fitted results aren't saved, so you re-click Fit & Forecast after loading.

Install

install.packages("remotes")
remotes::install_local("forecastsuite")

Run the app

forecastsuite::run_app()

In RStudio, the app is also available from the Addins menu ("Run forecastsuite") -- no need to type the command.

Learn more

vignette("forecastsuite") walks through the bundled app tab-by-tab and the model registry API for scripting forecasts directly in R, without the app.

Prophet support (optional)

install.packages("prophet")

Everything else works without prophet — list_models() simply omits Prophet until it's installed. prophet pulls in a compiled Stan backend (rstan/StanHeaders) that has a real history of install/check fragility on some platforms, so it's kept optional (Suggests, not Imports) rather than force-loaded by library(forecastsuite). Once installed, Prophet's changepoint/seasonality/holiday priors are exposed on the Model tab like any other model's parameters, saved/restored with a project file, and included in the "Show Code" panel.

LSTM support (optional)

install.packages("torch")
torch::install_torch()

Everything else works without torch — list_models() simply omits LSTM until it's installed. Once available, its epochs / hidden units / lookback window / learning rate are exposed on the Model tab like any other model's parameters, saved/restored with a project file, and included in the "Show Code" panel.

Documentation (contributors)

Every exported function has real roxygen2 comments in its R/ source, but NAMESPACE/man/*.Rd are regenerated from them, not hand-edited. After changing a function's documentation or @export status:

Rscript document.R

(installs roxygen2 if needed, then runs roxygen2::roxygenise()). Commit the updated NAMESPACE/man/.

Relationship to the hosted app

R/data_utils.R, R/metric_utils.R, R/holidays_helpers.R, and R/horizon_utils.R are adapted from the parent repo's helpers/ files (same behavior, package-namespaced). They are intentionally copied, not shared, so this package and the hosted app can evolve independently without one destabilizing the other. If you fix a bug in one copy, check whether the other needs the same fix.