Merge pull request #92 from databricks-industry-solutions/debug-sktime

added prophet via sktime

Author: ryuta-yoshimatsu

README.md

@@ -12,6 +12,7 @@ Get started now!
 
 ## What's New
 
+- Feb 2025: [Prophet](https://www.sktime.net/en/stable/api_reference/auto_generated/sktime.forecasting.fbprophet.Prophet.html) is available for univariate forecasting via `SKTimeProphet`. Try the [notebook](https://github.com/databricks-industry-solutions/many-model-forecasting/blob/main/examples/daily/local_univariate_daily).
 - Feb 2025: Added a post evaluation notebook that shows how to run fine-grained model selection after running MMF. Try the [notebook](https://github.com/databricks-industry-solutions/many-model-forecasting/blob/main/examples/post-evaluation-analysis.ipynb).
 - Jan 2025: [TimesFM](https://github.com/google-research/timesfm) is available for univariate and covariate forecasting. Try the notebooks: [univariate](https://github.com/databricks-industry-solutions/many-model-forecasting/blob/main/examples/daily/foundation_daily.py) and [covariate](https://github.com/databricks-industry-solutions/many-model-forecasting/blob/main/examples/external_regressors/foundation_external_regressors_daily.py).
 - Jan 2025: [Chronos Bolt](https://github.com/amazon-science/chronos-forecasting) models are available for univariate forecasting. Try the [notebook](https://github.com/databricks-industry-solutions/many-model-forecasting/blob/main/examples/daily/foundation_daily.py).
@@ -51,6 +52,7 @@ active_models = [
     "RFableNNETAR",
     "RFableEnsemble",
     "RDynamicHarmonicRegression",
+    "SKTimeProphet",
     "SKTimeTBats",
     "SKTimeLgbmDsDt",
 ]

examples/daily/local_univariate_daily.py

@@ -152,6 +152,7 @@ def transform_group(df):
     "RFableEnsemble",
     "RDynamicHarmonicRegression",
     "SKTimeTBats",
+    "SKTimeProphet",
     "SKTimeLgbmDsDt",
     ]
 

examples/external_regressors/local_univariate_external_regressors_daily.py

@@ -127,6 +127,7 @@
     "RFableEnsemble",
     "RDynamicHarmonicRegression",
     "SKTimeTBats",
+    "SKTimeProphet",
     "SKTimeLgbmDsDt",
 ]
 

examples/hourly/local_univariate_hourly.py

@@ -121,7 +121,7 @@ def transform_group(df):
 # MAGIC %md ### Models
 # MAGIC Let's configure a list of models we are going to apply to our time series for evaluation and forecasting. A comprehensive list of all supported models is available in [mmf_sa/models/models_conf.yaml](https://github.com/databricks-industry-solutions/many-model-forecasting/blob/main/mmf_sa/models/models_conf.yaml). Look for the models where `model_type: local`; these are the local models we import from [statsforecast](https://github.com/Nixtla/statsforecast). Check their documentations for the description of each model. 
 # MAGIC
-# MAGIC *Note that hourly forecasting is currently not supported for `r fable` and `sktime` models.*
+# MAGIC *Note that hourly forecasting is currently not supported for `r fable` models.*
 
 # COMMAND ----------
 
@@ -140,6 +140,9 @@ def transform_group(df):
     "StatsForecastCrostonClassic",
     "StatsForecastCrostonOptimized",
     "StatsForecastCrostonSBA",
+    "SKTimeTBats",
+    "SKTimeProphet",
+    "SKTimeLgbmDsDt",
     ]
 
 # COMMAND ----------

examples/m5/local_univariate_daily_m5.py

@@ -57,6 +57,7 @@
     "RFableEnsemble",
     "RDynamicHarmonicRegression",
     "SKTimeTBats",
+    "SKTimeProphet",
     "SKTimeLgbmDsDt",
 ]
 

examples/monthly/local_univariate_monthly.py

@@ -161,6 +161,7 @@ def transform_group(df):
     "RFableEnsemble",
     "RDynamicHarmonicRegression",
     "SKTimeTBats",
+    "SKTimeProphet",
     "SKTimeLgbmDsDt",
 ]
 

examples/weekly/local_univariate_weekly.py

@@ -150,6 +150,7 @@ def transform_group(df):
     "RFableEnsemble",
     "RDynamicHarmonicRegression",
     "SKTimeTBats",
+    "SKTimeProphet",
     "SKTimeLgbmDsDt",
     ]
 

mmf_sa/models/models_conf.yaml

@@ -166,27 +166,40 @@ models:
     model_spec:
       fourier_terms:
 
-  SKTimeLgbmDsDt:
+  SKTimeTBats:
     module: mmf_sa.models.sktime.SKTimeForecastingPipeline
-    model_class: SKTimeLgbmDsDt
+    model_class: SKTimeTBats
     framework: SKTime
     model_type: local
     enable_gcv: false
     model_spec:
-      deseasonalise_model: multiplicative
+      box_cox: True
+      use_trend: True
       season_length: 7
-      detrend_poly_degree: 2
 
-  SKTimeTBats:
+  SKTimeProphet:
     module: mmf_sa.models.sktime.SKTimeForecastingPipeline
-    model_class: SKTimeTBats
+    model_class: SKTimeProphet
     framework: SKTime
     model_type: local
     enable_gcv: false
     model_spec:
-      box_cox: True
-      use_trend: True
+      growth: linear
+      yearly_seasonality: auto
+      weekly_seasonality: auto
+      daily_seasonality: auto
+      seasonality_mode: additive
+
+  SKTimeLgbmDsDt:
+    module: mmf_sa.models.sktime.SKTimeForecastingPipeline
+    model_class: SKTimeLgbmDsDt
+    framework: SKTime
+    model_type: local
+    enable_gcv: false
+    model_spec:
+      deseasonalise_model: multiplicative
       season_length: 7
+      detrend_poly_degree: 2
 
   NeuralForecastRNN:
     module: mmf_sa.models.neuralforecast.NeuralForecastPipeline

mmf_sa/models/sktime/SKTimeForecastingPipeline.py

@@ -8,6 +8,7 @@
     ForecastingGridSearchCV,
 )
 from sktime.forecasting.tbats import TBATS
+from sktime.forecasting.fbprophet import Prophet
 from sktime.forecasting.compose import make_reduction
 from sktime.forecasting.compose import TransformedTargetForecaster
 from sktime.transformations.series.detrend import Detrender, ConditionalDeseasonalizer
@@ -47,7 +48,9 @@ def prepare_data(self, df: pd.DataFrame) -> pd.DataFrame:
         return df
 
     def fit(self, x, y=None):
-        if self.params.get("enable_gcv", False) and self.model is None and self.param_grid:
+        if (self.params.get("enable_gcv", False)
+                and self.model is None
+                and self.param_grid):
             _model = self.create_model()
             cv = SlidingWindowSplitter(
                 initial_window=int(len(x) - self.params.prediction_length * 4),
@@ -68,8 +71,8 @@ def predict(self, hist_df: pd.DataFrame, val_df: pd.DataFrame = None):
             ForecastingHorizon(np.arange(1, self.params.prediction_length + 1))
         )
         date_idx = pd.date_range(
-            _df.index.max().to_timestamp(freq=self.params.freq) + pd.DateOffset(days=1),
-            _df.index.max().to_timestamp(freq=self.params.freq) + pd.DateOffset(days=self.params.prediction_length),
+            _df.index.max().to_timestamp(freq=self.params.freq) + self.one_ts_offset,
+            _df.index.max().to_timestamp(freq=self.params.freq) + self.prediction_length_offset,
             freq=self.params.freq,
             name=self.params.date_col,
         )
@@ -82,6 +85,48 @@ def forecast(self, x, spark=None):
         return self.predict(x)
 
 
+class SKTimeTBats(SKTimeForecastingPipeline):
+    def __init__(self, params):
+        super().__init__(params)
+
+    def create_model(self) -> BaseForecaster:
+        model = TBATS(
+            sp=int(self.model_spec.get("season_length")),
+            use_trend=self.model_spec.get("use_trend"),
+            use_box_cox=self.model_spec.get("box_cox"),
+            n_jobs=-1,
+        )
+        return model
+
+    def create_param_grid(self):
+        return {
+            "use_trend": [True, False],
+            "use_box_cox": [True, False],
+            "sp": [1, 7, 14],
+        }
+
+class SKTimeProphet(SKTimeForecastingPipeline):
+    def __init__(self, params):
+        super().__init__(params)
+
+    def create_model(self) -> BaseForecaster:
+        model = Prophet(
+            freq=self.params.freq,
+            growth = self.model_spec.get("growth"),
+            yearly_seasonality=self.model_spec.get("yearly_seasonality"),
+            weekly_seasonality=self.model_spec.get("weekly_seasonality"),
+            daily_seasonality=self.model_spec.get("daily_seasonality"),
+            seasonality_mode=self.model_spec.get("seasonality_mode"),
+        )
+        return model
+
+    def create_param_grid(self):
+        return {
+            "growth": ['linear', 'logarithmic'],
+            "seasonality_mode": ['additive', 'multiplicative'],
+        }
+
+
 class SKTimeLgbmDsDt(SKTimeForecastingPipeline):
     def __init__(self, params):
         super().__init__(params)
@@ -126,24 +171,3 @@ def create_param_grid(self):
                 self.params.prediction_length * 2,
             ],
         }
-
-
-class SKTimeTBats(SKTimeForecastingPipeline):
-    def __init__(self, params):
-        super().__init__(params)
-
-    def create_model(self) -> BaseForecaster:
-        model = TBATS(
-            sp=int(self.model_spec.get("season_length")),
-            use_trend=self.model_spec.get("use_trend"),
-            use_box_cox=self.model_spec.get("box_cox"),
-            n_jobs=-1,
-        )
-        return model
-
-    def create_param_grid(self):
-        return {
-            "use_trend": [True, False],
-            "use_box_cox": [True, False],
-            "sp": [1, 7, 14],
-        }