New parameter added in order to replace negative predictions with 0

python/prophet/forecaster.py

@@ -25,6 +25,7 @@
 logger.setLevel(logging.INFO)
 NANOSECONDS_TO_SECONDS = 1000 * 1000 * 1000
 
+
 class Prophet(object):
     """Prophet forecaster.
 
@@ -77,6 +78,7 @@ class Prophet(object):
     stan_backend: str as defined in StanBackendEnum default: None - will try to
         iterate over all available backends and find the working one
     holidays_mode: 'additive' or 'multiplicative'. Defaults to seasonality_mode.
+    negative_prediction_values: bool check to set all negative prediction values in the DataFrame to 0.
     """
 
     def __init__(
@@ -99,6 +101,7 @@ def __init__(
             stan_backend=None,
             scaling: str = 'absmax',
             holidays_mode=None,
+            negative_prediction_values=True
     ):
         self.growth = growth
 
@@ -150,6 +153,7 @@ def __init__(
         self.train_component_cols = None
         self.component_modes = None
         self.train_holiday_names = None
+        self.negative_prediction_values = negative_prediction_values
         self.fit_kwargs = {}
         self.validate_inputs()
         self._load_stan_backend(stan_backend)
@@ -1185,6 +1189,37 @@ def calculate_initial_params(self, num_total_regressors: int) -> ModelParams:
             sigma_obs=1.0,
         )
 
+    def calculate_and_clip_percentile(self, data, component, comp, lower_p, upper_p):
+        """
+        A helper function to calculate the lower and upper percentiles for a given component.
+
+        Parameters:
+        - data: dict or similar
+            The data structure where the calculated percentiles will be stored.
+        - component: str
+            The name of the component for which the percentiles are being calculated.
+        - comp: array-like
+            The data for which percentiles are to be calculated.
+        - lower_p: float
+            The percentile to calculate for the lower bound.
+        - upper_p: float
+            The percentile to calculate for the upper bound.
+
+        Returns:
+        - None
+            The function directly modifies the `data` structure by adding the lower and upper percentile values.
+        """
+        lower = self.percentile(comp, lower_p, axis=1)
+        upper = self.percentile(comp, upper_p, axis=1)
+
+        if not self.negative_prediction_values:
+            lower = np.clip(lower, a_min=0, a_max=None)
+            upper = np.clip(upper, a_min=0, a_max=None)
+
+        data[component + '_lower'] = lower
+        data[component + '_upper'] = upper
+
+
     def fit(self, df, **kwargs):
         """Fit the Prophet model.
 
@@ -1282,6 +1317,9 @@ def predict(self, df: pd.DataFrame = None, vectorized: bool = True) -> pd.DataFr
             cols.append('cap')
         if self.logistic_floor:
             cols.append('floor')
+        if not self.negative_prediction_values:
+            df['trend'] = df['trend'].clip(lower=0)
+
         # Add in forecast components
         df2 = pd.concat((df[cols], intervals, seasonal_components), axis=1)
         df2['yhat'] = (
@@ -1415,14 +1453,17 @@ def predict_seasonal_components(self, df):
             comp = np.matmul(X, beta_c.transpose())
             if component in self.component_modes['additive']:
                 comp *= self.y_scale
-            data[component] = np.nanmean(comp, axis=1)
+
+            if self.negative_prediction_values:
+                data[component] = np.nanmean(comp, axis=1)
+            else:
+                data[component] = np.clip(np.nanmean(comp, axis=1), a_min=0, a_max=None)
+
             if self.uncertainty_samples:
-                data[component + '_lower'] = self.percentile(
-                    comp, lower_p, axis=1,
-                )
-                data[component + '_upper'] = self.percentile(
-                    comp, upper_p, axis=1,
+                self.calculate_and_clip_percentile(
+                    data, component, comp, lower_p, upper_p
                 )
+
         return pd.DataFrame(data)
 
     def predict_uncertainty(self, df: pd.DataFrame, vectorized: bool) -> pd.DataFrame:
@@ -1444,10 +1485,9 @@ def predict_uncertainty(self, df: pd.DataFrame, vectorized: bool) -> pd.DataFram
 
         series = {}
         for key in ['yhat', 'trend']:
-            series['{}_lower'.format(key)] = self.percentile(
-                sim_values[key], lower_p, axis=1)
-            series['{}_upper'.format(key)] = self.percentile(
-                sim_values[key], upper_p, axis=1)
+            self.calculate_and_clip_percentile(
+                series, key, sim_values[key], lower_p, upper_p
+            )
 
         return pd.DataFrame(series)
 

python/prophet/tests/test_prophet.py

@@ -255,6 +255,7 @@ def test_make_future_dataframe_include_history(self, daily_univariate_ts, backen
 
         assert len(future) == train.shape[0] + 3
 
+
 class TestProphetTrendComponent:
     def test_invalid_growth_input(self, backend):
         msg = 'Parameter "growth" should be "linear", ' '"logistic" or "flat".'
@@ -433,6 +434,22 @@ def test_override_n_changepoints(self, daily_univariate_ts, backend):
         cp = m.changepoints_t
         assert cp.shape[0] == 15
 
+    @pytest.mark.parametrize(
+        "expected",
+        [5.656087514685135],
+    )
+    def test_without_negative_predictions(self, subdaily_univariate_ts, backend, expected):
+        test_days = 280
+        train, test = train_test_split(subdaily_univariate_ts, test_days)
+        forecaster = Prophet(stan_backend=backend, negative_prediction_values=False)
+        forecaster.fit(train, seed=1237861298)
+        np.random.seed(876543987)
+        future = forecaster.make_future_dataframe(test_days, include_history=False)
+        future = forecaster.predict(future)
+        res = rmse(future["yhat"], test["y"])
+        tolerance = 1e-5
+        assert res == pytest.approx(expected, rel=tolerance), "backend: {}".format(forecaster.stan_backend)
+
 
 class TestProphetSeasonalComponent:
     def test_fourier_series_weekly(self, daily_univariate_ts):
@@ -877,7 +894,6 @@ def test_subdaily_holidays(self, subdaily_univariate_ts, backend):
         assert sum(fcst["special_day"] == 0) == 575
 
 
-
 class TestProphetRegressors:
     def test_added_regressors(self, daily_univariate_ts, backend):
         m = Prophet(stan_backend=backend)