fix: fix filtering of NaNs in Prophet preprocessing (#219)

crates/augurs-prophet/src/data.rs

@@ -132,6 +132,52 @@ impl TrainingData {
         Ok(self)
     }
 
+    /// Remove any NaN values from the `y` column, and the corresponding values
+    /// in the other columns.
+    ///
+    /// This handles updating all columns and `n` appropriately.
+    ///
+    /// NaN values in other columns are retained.
+    pub(crate) fn filter_nans(mut self) -> Self {
+        let mut n = self.n;
+        let mut keep = vec![true; self.n];
+        self.y = self
+            .y
+            .into_iter()
+            .zip(keep.iter_mut())
+            .filter_map(|(y, keep)| {
+                if y.is_nan() {
+                    *keep = false;
+                    n -= 1;
+                    None
+                } else {
+                    Some(y)
+                }
+            })
+            .collect();
+
+        fn retain<T>(v: &mut Vec<T>, keep: &[bool]) {
+            let mut iter = keep.iter();
+            v.retain(|_| *iter.next().unwrap());
+        }
+
+        self.n = n;
+        retain(&mut self.ds, &keep);
+        if let Some(cap) = self.cap.as_mut() {
+            retain(cap, &keep);
+        }
+        if let Some(floor) = self.floor.as_mut() {
+            retain(floor, &keep);
+        }
+        for v in self.x.values_mut() {
+            retain(v, &keep);
+        }
+        for v in self.seasonality_conditions.values_mut() {
+            retain(v, &keep);
+        }
+        self
+    }
+
     #[cfg(test)]
     pub(crate) fn head(mut self, n: usize) -> Self {
         self.n = n;
@@ -298,3 +344,19 @@ impl PredictionData {
         Ok(self)
     }
 }
+
+#[cfg(test)]
+mod test {
+    use crate::testdata::daily_univariate_ts;
+
+    #[test]
+    fn filter_nans() {
+        let mut data = daily_univariate_ts();
+        let expected_len = data.n - 1;
+        data.y[10] = f64::NAN;
+        let data = data.filter_nans();
+        assert_eq!(data.n, expected_len);
+        assert_eq!(data.y.len(), expected_len);
+        assert_eq!(data.ds.len(), expected_len);
+    }
+}

crates/augurs-prophet/src/forecaster.rs

@@ -151,7 +151,7 @@ impl Predict for FittedProphetForecaster {
     }
 }
 
-#[cfg(test)]
+#[cfg(all(test, feature = "wasmstan"))]
 mod test {
 
     use augurs_core::{Fit, Predict};

crates/augurs-prophet/src/prophet.rs

@@ -863,4 +863,16 @@ mod test_fit {
             &[0.781831, 0.623490, 0.974928, -0.222521, 0.433884, -0.900969],
         );
     }
+
+    // Regression test for https://github.com/grafana/augurs/issues/209.
+    #[test]
+    fn fit_with_nans() {
+        let test_days = 30;
+        let (mut train, _) = train_test_splitn(daily_univariate_ts(), test_days);
+        train.y[10] = f64::NAN;
+        let opt = MockOptimizer::new();
+        let mut prophet = Prophet::new(Default::default(), opt);
+        // Should not panic.
+        prophet.fit(train.clone(), Default::default()).unwrap();
+    }
 }

crates/augurs-prophet/src/prophet/prep.rs

@@ -194,7 +194,7 @@ pub(super) struct Features {
 }
 
 impl<O> Prophet<O> {
-    pub(super) fn preprocess(&mut self, mut data: TrainingData) -> Result<Preprocessed, Error> {
+    pub(super) fn preprocess(&mut self, data: TrainingData) -> Result<Preprocessed, Error> {
         let n = data.ds.len();
         if n != data.y.len() {
             return Err(Error::MismatchedLengths {
@@ -207,12 +207,7 @@ impl<O> Prophet<O> {
         if n < 2 {
             return Err(Error::NotEnoughData);
         }
-        (data.ds, data.y) = data
-            .ds
-            .into_iter()
-            .zip(data.y)
-            .filter(|(_, y)| !y.is_nan())
-            .unzip();
+        let data = data.filter_nans();
 
         let mut history_dates = data.ds.clone();
         history_dates.sort_unstable();