Fix: R ARIMAX wrapper when features are not provided

src/gluonts/ext/r_forecast/R/univariate_forecast_methods.R

@@ -48,6 +48,49 @@ arima <- function(ts, params) {
 
 
 fourier.arima.xreg <- function(ts, params, xreg_in, xreg_out){
+
+    if (missing(xreg_in)){
+        fourier.arima(ts, params)
+    } else {
+
+        fourier.frequency.low.periods <- 4
+        fourier.ratio.threshold.low.periods <- 18
+        fourier.frequency.high.periods <- 52
+        fourier.ratio.threshold.high.periods <- 2
+        fourier.order <- 4
+
+        period <- frequency(ts)
+        len_ts <- length(ts)
+        fourier_ratio <- len_ts / period
+
+        fourier <- FALSE
+
+        if ((period > fourier.frequency.low.periods
+            && fourier_ratio > fourier.ratio.threshold.low.periods)
+            || (period >= fourier.frequency.high.periods
+            && fourier_ratio > fourier.ratio.threshold.high.periods)) {
+            # When the period is high, auto.arima becomes unstable
+            # per Rob's suggestion, we use Fourier series instead
+            # cf. https://robjhyndman.com/hyndsight/longseasonality/
+            fourier <- TRUE
+        }
+
+        if (fourier == TRUE) {
+            K <- min(fourier.order, floor(frequency(ts) / 2))
+            seasonal <- FALSE
+            xreg <- forecast::fourier(ts, K=K)
+            xreg_in <- as.matrix(xreg_in, xreg)
+            model <- forecast::auto.arima(ts, seasonal = seasonal, xreg = xreg_in, trace=TRUE)
+
+            xreg <- forecast::fourier(ts, K=K, h=params$prediction_length)
+            xreg_out <- as.matrix(xreg_out, xreg)
+
+            handleModel(model, params, xreg_out)
+        } else {
+            model <- forecast::auto.arima(ts, xreg = xreg_in, trace=TRUE)
+            handleModel(model, params, xreg_out)
+        }
+
     fourier.frequency.low.periods <- 4
     fourier.ratio.threshold.low.periods <- 18
     fourier.frequency.high.periods <- 52

src/gluonts/ext/r_forecast/_univariate_predictor.py

@@ -145,7 +145,6 @@ def _get_r_forecast(self, data: Dict) -> Dict:
             import rpy2.robjects.numpy2ri
 
             rpy2.robjects.numpy2ri.activate()
-
             data["feat_dynamic_real"] = np.transpose(data["feat_dynamic_real"])
             nrow, ncol = data["feat_dynamic_real"].shape
             xreg_in = self._robjects.r.matrix(

test/ext/r_forecast/test_r_multi_seasonality.py

@@ -146,8 +146,7 @@ def test_compare_arimas():
 
     assert fourier_arima_eval_metrics["MASE"] < arima_eval_metrics["MASE"]
 
-
-## Below shows improvement in metric when proper x_regressors are included #
+## Below shows improvement in metric when proper x_regressors are included #   
 
 dataset_xreg = [
     {