CHANGES IN NNFOR VERSION 0.10.0

• New torch backend for mlp(): backend = "torch" (torch is optional, in Suggests). By default it reproduces neuralnet’s training setup (logistic hidden layers, Rprop, the same stopping rule and starting weights drawn as neuralnet draws them). Activation functions (act.fct) and optimizers (algorithm) can be switched, using neuralnet’s or torch’s names. All ensemble members are trained simultaneously as one batch of tensors.
• Custom architectures with torch: hd can be a function returning the hidden layers as a torch module; nnfor handles inputs, output layer and forecasting.
• New compare_backends(): fits and forecasts a series with each backend, reporting run times and in-sample and out-of-sample MSE.
• ?mlp documents the torch settings, suggested settings for torch and the differences between the backends.
• nnfor_map() shows the backend used by a fitted network.
• The hidden-node search (hd.auto.type = "valid" or "cv") now uses the training settings passed in ..., so the size is chosen for the network that is then trained. Previously it used default settings.
• A warning is given when hd.auto.type is set but ignored because hd is given.

CHANGES IN NNFOR VERSION 0.9.9

• Major refactoring: Preprocessing split into focused steps (differencing and scaling, lagged inputs, seasonal dummies, input selection). Training and forecasting build inputs with the same function, so they cannot disagree.
• Network training goes through a backend interface (neuralnet for now) to prepare for further backends, such as torch.
• Plotting consolidated. Plots no longer override user graphical settings (par(), layouts, colours) and default line colours follow par("fg").
• New:nnfor_map() shows how information flows through the package when fitting (stage = "fit") or forecasting (stage = "forecast") with mlp() or elm(). Plot it for a diagram or print it for an outline. Given a fitted network, the map shows the values that fit used and greys out unused steps.
• m when used, it overrides the seasonal periods of y (vector for multiple seasonalities). Its default is now NULL. The periods are stored in the fitted object as ff.
• Plots accept ... for the plot frame and *.args lists for each drawn component. Time series plots accept add = TRUE.
• plot(fit, which = "fit") reproduces the outplot figure after fitting. plot_network() and plot_fit() are exported.
• plot(forecast) can show the in-sample fit (fitted = TRUE). Default y-limits follow xlim.
• Fitted objects store the fit of each ensemble member (fitted.all).
• Ensemble members can be trained in parallel with options(nnfor.cores = ), giving an integer or a cluster from parallel::makeCluster(). This covers MLP training, the hidden-node search, ELM output weights and elm.fast(). Random numbers are drawn before members are dispatched, so results are identical for any number of cores and to sequential runs.
• Requires tsutils >= 0.9.5, whose seasdummy() fixes binary seasonal dummies that were out of phase for the last observations of the sample when it did not start at the first season. Fits with binary dummies change as a result.
• When input selection rejects every univariate lag, but regressors or dummies remain, the network uses those only; previously all lags were kept.
• elm.fast() logistic output applies the logistic function (the inverse link of the binomial model the output weights are estimated with) and reverses the stored target scaling, as linear output does. Previously the hidden-layer activation was used and predictions were rescaled by their own range (a single prediction gave NaN).
• Reusing a model (model =) copies its specification exactly, rather than re-detecting differencing and seasonality on the new data.
• mlp() failed with “argument is of length zero” when only some ensemble members converged (a neuralnet issue). Members that do not converge are now dropped, with a warning, as intended.
• Dummies rebuilt after input selection used the wrong seasonal periods for multiple seasonalities.
• mlp.thief() and elm.thief() passed xreg as y to forecast().
• elm(barebone = TRUE, type = "lm") failed; "lm" and "ls" are now synonyms.
• ELM with several hidden layers used wrong weights.
• ELM network plot greyed out the wrong hidden nodes.
• One-step forecasts with trigonometric dummies failed.
• Clear errors when forecasting without required xreg, when no network converges, and for barebone = TRUE with several hidden layers.

CHANGED IN NNFOR VERSION 0.9.7 (07 JUL 2022)

• Added warning when predict() is called for ELM and MLP.

CHANGED IN NNFOR VERSION 0.9.7 (05 JUL 2022)

• Updated dependencies

CHANGES IN NNFOR VERSION 0.9.6 (15 JAN 2019)

• Updated dependencies

CHANGES IN NNFOR VERSION 0.9.5 (12 JAN 2019)

• Removed functions mseastest, cmav and lagmatrix from nnfor. These are now available in package tsutils.
• Updated license to GPL-3.

CHANGES IN NNFOR VERSION 0.9.4 (11 JUL 2018)

• Fix for msts time series and allow.det.season argument.

CHANGES IN NNFOR VERSION 0.9.3 (01 Apr 2018)

• Changes in how trigonometric seasonal dummies are handled.

CHANGES IN NNFOR VERSION 0.9.2 (11 Dec 2017)

• Documentation fixes.

CHANGES IN NNFOR VERSION 0.9.2 (30 Nov 2017)

• Added option to use previously specified mlp and elm (see arguments model and retrain).

CHANGES IN NNFOR VERSION 0.9.1 (29 Oct 2017)

• Fix on univariate keep vector when no univariate lags are selected (lags=0).