diff --git a/xesn/driver.py b/xesn/driver.py
index 98cda83..b01ed12 100644
--- a/xesn/driver.py
+++ b/xesn/driver.py
@@ -209,6 +209,8 @@ def run_test(self):
                     n_steps=self.config["testing"]["n_steps"],
                     n_spinup=self.config["testing"]["n_spinup"]
                     )
+            xds["prediction"] = data.normalize_inverse(xds["prediction"], keep_attrs=True)
+            xds["truth"] = data.normalize_inverse(xds["truth"], keep_attrs=True)
             xds.to_zarr(join(self.output_directory, f"test-{i}.zarr"), mode="w")
 
         self.localtime.stop()
diff --git a/xesn/test/xdata.py b/xesn/test/xdata.py
index 4a9270c..28c65f2 100644
--- a/xesn/test/xdata.py
+++ b/xesn/test/xdata.py
@@ -67,7 +67,9 @@ def test_data():
                 "z": np.linspace(  0, 50, 5),
                 "time": np.linspace(0, 2000, 200),
                 },
-            dims=tester.dimensions)
+            dims=tester.dimensions,
+            attrs={"description": "This is some test data!"},
+            )
     u.name = tester.field_name
     u.to_dataset().to_zarr(tester.zstore_path, mode="w")
     yield u
@@ -166,16 +168,39 @@ def test_subsample_type(self, test_data):
             xd.subsample(test_data, mode="training")
 
 
-    def test_normalize(self, test_data):
+    @pytest.mark.parametrize(
+            "keep_attrs", (True, False)
+        )
+    def test_normalize(self, test_data, keep_attrs):
         xd = XData(self.field_name,
                    self.zstore_path,
                    dimensions=self.dimensions,
                    normalization=self.normalization)
 
-        test = xd.normalize(test_data)
+        test = xd.normalize(test_data, keep_attrs=keep_attrs)
         expected = (test_data - self.normalization["bias"]) / self.normalization["scale"]
         assert_allclose(test, expected)
 
+        if keep_attrs:
+            assert test.attrs == test_data.attrs
+
+
+    @pytest.mark.parametrize(
+            "keep_attrs", (True, False)
+        )
+    def test_normalize_inverse(self, test_data, keep_attrs):
+        xd = XData(self.field_name,
+                   self.zstore_path,
+                   dimensions=self.dimensions,
+                   normalization=self.normalization)
+
+        test = xd.normalize_inverse(test_data, keep_attrs)
+        expected = test_data *  self.normalization["scale"] + self.normalization["bias"]
+        assert_allclose(test, expected)
+
+        if keep_attrs:
+            assert test.attrs == test_data.attrs
+
 
     # Some repetition here, but I think it's worth it
     @pytest.mark.parametrize(
diff --git a/xesn/xdata.py b/xesn/xdata.py
index a68d626..317386d 100644
--- a/xesn/xdata.py
+++ b/xesn/xdata.py
@@ -99,7 +99,7 @@ def subsample(self, xda, mode):
         return xda
 
 
-    def normalize(self, xda):
+    def normalize(self, xda, keep_attrs=False):
         """Very simple, this may eventually be hooked up with scikit learn for more advanced normalization.
         Right now, normalize with scalars :attr:`bias` and :attr:`scale` as
 
@@ -108,18 +108,44 @@ def normalize(self, xda):
 
         Args:
             xda (xarray.DataArray): with the field to be normalized
+            keep_attrs (bool): if True, keep attributes in xda
 
         Returns:
             xda (xarray.DataArray): normalized
         """
 
-        # good practice?
         if self.normalization is None:
             return xda
 
-        # TODO: not really clear how to handle other cases...
-        # is it worth handling what options should be here?
-        # assert self.normalization["type"] == "normal-scalar"
         bias = self.normalization.get("bias", 0.)
         scale= self.normalization.get("scale", 1.)
-        return (xda - bias)/scale
+
+        with xr.set_options(keep_attrs=keep_attrs):
+            result = (xda - bias)/scale
+
+        return result
+
+
+    def normalize_inverse(self, xda, keep_attrs=False):
+        """Do the inverse operation of :meth:`normalize`, i.e.,
+
+        .. math::
+            xda * scale + bias
+
+        Args:
+            xda (xarray.DataArray): with the normalized field to be re scaled
+            keep_attrs (bool): if True, keep attributes in xda
+
+        Returns:
+            xda (xarray.DataArray): scaled and biased like original data
+        """
+        if self.normalization is None:
+            return xda
+
+        bias = self.normalization.get("bias", 0.)
+        scale= self.normalization.get("scale", 1.)
+
+        with xr.set_options(keep_attrs=keep_attrs):
+            result = xda*scale + bias
+
+        return result