autogluon.timeseries 1.0.1b20240327__tar.gz → 1.0.1b20240403__tar.gz

{autogluon.timeseries-1.0.1b20240327 → autogluon.timeseries-1.0.1b20240403}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: autogluon.timeseries
-Version: 1.0.1b20240327
+Version: 1.0.1b20240403
 Summary: AutoML for Image, Text, and Tabular Data
 Home-page: https://github.com/autogluon/autogluon
 Author: AutoGluon Community
@@ -26,7 +26,7 @@ Description:
         [![Continuous Integration](https://github.com/autogluon/autogluon/actions/workflows/continuous_integration.yml/badge.svg)](https://github.com/autogluon/autogluon/actions/workflows/continuous_integration.yml)
         [![Platform Tests](https://github.com/autogluon/autogluon/actions/workflows/platform_tests-command.yml/badge.svg?event=schedule)](https://github.com/autogluon/autogluon/actions/workflows/platform_tests-command.yml)
         
-        [Install Instructions](https://auto.gluon.ai/stable/install.html) | [Documentation](https://auto.gluon.ai/stable/index.html) | [Release Notes](https://auto.gluon.ai/stable/whats_new/index.html)
+        [Installation](https://auto.gluon.ai/stable/install.html) | [Documentation](https://auto.gluon.ai/stable/index.html) | [Release Notes](https://auto.gluon.ai/stable/whats_new/index.html)
         
         AutoGluon automates machine learning tasks enabling you to easily achieve strong predictive performance in your applications.  With just a few lines of code, you can train and deploy high-accuracy machine learning and deep learning models on image, text, time series, and tabular data.
         </div>

{autogluon.timeseries-1.0.1b20240327 → autogluon.timeseries-1.0.1b20240403}/src/autogluon/timeseries/dataset/ts_dataframe.py

@@ -765,11 +765,19 @@ class TimeSeriesDataFrame(pd.DataFrame, TimeSeriesDataFrameDeprecatedMixin):
                 "(for example, using the `convert_frequency` method)."
             )
 
-        grouped_df = pd.DataFrame(self).groupby(level=ITEMID, sort=False, group_keys=False)
+        # Convert to pd.DataFrame for faster processing
+        df = pd.DataFrame(self)
+
+        # Skip filling if there are no NaNs
+        if not df.isna().any(axis=None):
+            return self
+
+        grouped_df = df.groupby(level=ITEMID, sort=False, group_keys=False)
         if method == "auto":
             filled_df = grouped_df.ffill()
-            # Fill missing values at the start of each time series with bfill
-            filled_df = filled_df.groupby(level=ITEMID, sort=False, group_keys=False).bfill()
+            # If necessary, fill missing values at the start of each time series with bfill
+            if filled_df.isna().any(axis=None):
+                filled_df = filled_df.groupby(level=ITEMID, sort=False, group_keys=False).bfill()
         elif method in ["ffill", "pad"]:
             filled_df = grouped_df.ffill()
         elif method in ["bfill", "backfill"]:

{autogluon.timeseries-1.0.1b20240327 → autogluon.timeseries-1.0.1b20240403}/src/autogluon/timeseries/learner.py

@@ -1,7 +1,7 @@
 import logging
 import reprlib
 import time
-from typing import Any, Dict, List, Optional, Type, Union
+from typing import Any, Dict, List, Literal, Optional, Type, Union
 
 import pandas as pd
 
@@ -228,5 +228,32 @@ class TimeSeriesLearner(AbstractLearner):
         learner_info.pop("random_state", None)
         return learner_info
 
+    def persist_trainer(
+        self, models: Union[Literal["all", "best"], List[str]] = "all", with_ancestors: bool = False
+    ) -> List[str]:
+        """Loads models and trainer in memory so that they don't have to be
+        loaded during predictions
+
+        Returns
+        -------
+        list_of_models : List[str]
+            List of models persisted in memory
+        """
+        self.trainer = self.load_trainer()
+        return self.trainer.persist(models, with_ancestors=with_ancestors)
+
+    def unpersist_trainer(self) -> List[str]:
+        """Unloads models and trainer from memory. Models will have to be reloaded from disk
+        when predicting.
+
+        Returns
+        -------
+        list_of_models : List[str]
+            List of models removed from memory
+        """
+        unpersisted_models = self.load_trainer().unpersist()
+        self.trainer = None
+        return unpersisted_models
+
     def refit_full(self, model: str = "all") -> Dict[str, str]:
         return self.load_trainer().refit_full(model=model)

{autogluon.timeseries-1.0.1b20240327 → autogluon.timeseries-1.0.1b20240403}/src/autogluon/timeseries/models/abstract/abstract_timeseries_model.py

@@ -201,7 +201,9 @@ class AbstractTimeSeriesModel(AbstractModel):
         }
         return info
 
-    def fit(self, **kwargs) -> "AbstractTimeSeriesModel":
+    def fit(
+        self, train_data: TimeSeriesDataFrame, val_data: Optional[TimeSeriesDataFrame] = None, **kwargs
+    ) -> "AbstractTimeSeriesModel":
         """Fit timeseries model.
 
         Models should not override the `fit` method, but instead override the `_fit` method which
@@ -235,7 +237,10 @@ class AbstractTimeSeriesModel(AbstractModel):
         model: AbstractTimeSeriesModel
             The fitted model object
         """
-        return super().fit(**kwargs)
+        train_data = self.preprocess(train_data, is_train=True)
+        if self._get_tags()["can_use_val_data"] and val_data is not None:
+            val_data = self.preprocess(val_data, is_train=False)
+        return super().fit(train_data=train_data, val_data=val_data, **kwargs)
 
     def _fit(
         self,
@@ -290,6 +295,7 @@ class AbstractTimeSeriesModel(AbstractModel):
             data is given as a separate forecast item in the dictionary, keyed by the `item_id`s
             of input items.
         """
+        data = self.preprocess(data, is_train=False)
         predictions = self._predict(data=data, known_covariates=known_covariates, **kwargs)
         logger.debug(f"Predicting with model {self.name}")
         # "0.5" might be missing from the quantiles if self is a wrapper (MultiWindowBacktestingModel or ensemble)
@@ -415,6 +421,13 @@ class AbstractTimeSeriesModel(AbstractModel):
         hpo_executor.register_resources(self, k_fold=1, **kwargs)
         return self._hyperparameter_tune(hpo_executor=hpo_executor, **kwargs)
 
+    def persist(self) -> "AbstractTimeSeriesModel":
+        """Ask the model to persist its assets in memory, i.e., to predict with low latency. In practice
+        this is used for pretrained models that have to lazy-load model parameters to device memory at
+        prediction time.
+        """
+        return self
+
     def _hyperparameter_tune(
         self,
         train_data: TimeSeriesDataFrame,
@@ -481,7 +494,7 @@ class AbstractTimeSeriesModel(AbstractModel):
 
         return hpo_models, analysis
 
-    def preprocess(self, data: Any, **kwargs) -> Any:
+    def preprocess(self, data: TimeSeriesDataFrame, is_train: bool = False, **kwargs) -> Any:
         return data
 
     def get_memory_size(self, **kwargs) -> Optional[int]:
@@ -499,3 +512,20 @@ class AbstractTimeSeriesModel(AbstractModel):
             return {}
         else:
             return self._user_params.copy()
+
+    def _more_tags(self) -> dict:
+        """Encode model properties using tags, similar to sklearn & autogluon.tabular.
+
+        For more details, see `autogluon.core.models.abstract.AbstractModel._get_tags()` and https://scikit-learn.org/stable/_sources/developers/develop.rst.txt.
+
+        List of currently supported tags:
+        - allow_nan: Can the model handle data with missing values represented by np.nan?
+        - can_refit_full: Does it make sense to retrain the model without validation data?
+            See `autogluon.core.models.abstract._tags._DEFAULT_TAGS` for more details.
+        - can_use_val_data: Can model use val_data if it's provided to model.fit()?
+        """
+        return {
+            "allow_nan": False,
+            "can_refit_full": False,
+            "can_use_val_data": False,
+        }

{autogluon.timeseries-1.0.1b20240327 → autogluon.timeseries-1.0.1b20240403}/src/autogluon/timeseries/models/autogluon_tabular/mlforecast.py

@@ -85,6 +85,21 @@ class AbstractMLForecastModel(AbstractTimeSeriesModel):
         self._scaler: Optional[BaseTargetTransform] = None
         self._residuals_std_per_item: Optional[pd.Series] = None
         self._avg_residuals_std: Optional[float] = None
+        self._train_target_median: Optional[float] = None
+
+    def preprocess(self, data: TimeSeriesDataFrame, is_train: bool = False, **kwargs) -> Any:
+        if is_train:
+            # All-NaN series are removed; partially-NaN series in train_data are handled inside _generate_train_val_dfs
+            all_nan_items = data.item_ids[data[self.target].isna().groupby(ITEMID, sort=False).all()]
+            if len(all_nan_items):
+                data = data.query("item_id not in @all_nan_items")
+            return data
+        else:
+            data = data.fill_missing_values()
+            # Fill time series consisting of all NaNs with the median of target in train_data
+            if data.isna().any(axis=None):
+                data[self.target] = data[self.target].fillna(value=self._train_target_median)
+            return data
 
     def _get_extra_tabular_init_kwargs(self) -> dict:
         raise NotImplementedError
@@ -98,8 +113,6 @@ class AbstractMLForecastModel(AbstractTimeSeriesModel):
         return model_params
 
     def _get_mlforecast_init_args(self, train_data: TimeSeriesDataFrame, model_params: dict) -> dict:
-        # TODO: Support lag generation for all pandas frequencies
-        # TODO: Support date_feature generation for all pandas frequencies
         from mlforecast.target_transforms import Differences
 
         from .utils import MeanAbsScaler, StandardScaler
@@ -181,6 +194,10 @@ class AbstractMLForecastModel(AbstractTimeSeriesModel):
             items_to_keep = data.item_ids.to_series().sample(n=int(max_num_items))  # noqa: F841
             data = data.query("item_id in @items_to_keep")
 
+        # MLForecast.preprocess does not support missing values, but we will exclude them later from the training set
+        missing_entries = data.index[data[self.target].isna()]
+        data = data.fill_missing_values()
+
         num_items = data.num_items
         mlforecast_df = self._to_mlforecast_df(data, data.static_features)
 
@@ -197,6 +214,10 @@ class AbstractMLForecastModel(AbstractTimeSeriesModel):
 
         df = self._mask_df(df)
 
+        # We remove originally missing values filled via imputation from the training set
+        if len(missing_entries):
+            df = df.set_index(["unique_id", "ds"]).drop(missing_entries, errors="ignore").reset_index()
+
         if max_num_samples is not None and len(df) > max_num_samples:
             df = df.sample(n=max_num_samples)
 
@@ -246,6 +267,7 @@ class AbstractMLForecastModel(AbstractTimeSeriesModel):
 
         self._check_fit_params()
         fit_start_time = time.time()
+        self._train_target_median = train_data[self.target].median()
         # TabularEstimator is passed to MLForecast later to include tuning_data
         model_params = self._get_model_params()
 
@@ -355,7 +377,7 @@ class AbstractMLForecastModel(AbstractTimeSeriesModel):
         return predictions
 
     def _more_tags(self) -> dict:
-        return {"can_refit_full": True}
+        return {"allow_nan": True, "can_refit_full": True}
 
 
 class DirectTabularModel(AbstractMLForecastModel):

{autogluon.timeseries-1.0.1b20240327 → autogluon.timeseries-1.0.1b20240403}/src/autogluon/timeseries/models/chronos/model.py

@@ -18,11 +18,29 @@ logger = logging.getLogger(__name__)
 MODEL_CONFIGS = {
     "amazon/chronos-t5-tiny": {
         "num_gpus": 0,  # minimum number of required GPUs
+        "default_torch_dtype": "auto",
+        "default_batch_size": 16,
+    },
+    "amazon/chronos-t5-mini": {
+        "num_gpus": 0,
+        "default_torch_dtype": "auto",
+        "default_batch_size": 16,
+    },
+    "amazon/chronos-t5-small": {
+        "num_gpus": 1,
+        "default_torch_dtype": "bfloat16",
+        "default_batch_size": 16,
+    },
+    "amazon/chronos-t5-base": {
+        "num_gpus": 1,
+        "default_torch_dtype": "bfloat16",
+        "default_batch_size": 16,
+    },
+    "amazon/chronos-t5-large": {
+        "num_gpus": 1,
+        "default_torch_dtype": "bfloat16",
+        "default_batch_size": 8,
     },
-    "amazon/chronos-t5-mini": {"num_gpus": 0},
-    "amazon/chronos-t5-small": {"num_gpus": 1},
-    "amazon/chronos-t5-base": {"num_gpus": 1},
-    "amazon/chronos-t5-large": {"num_gpus": 1},
 }
 
 
@@ -124,7 +142,6 @@ class ChronosModel(AbstractTimeSeriesModel):
 
     # default number of samples for prediction
     default_num_samples: int = 20
-    default_batch_size: int = 16
     default_model_path = "amazon/chronos-t5-small"
     maximum_context_length = 512
 
@@ -149,7 +166,7 @@ class ChronosModel(AbstractTimeSeriesModel):
         self.device = hyperparameters.get("device")
 
         # if the model requires a GPU, set the torch dtype to bfloat16
-        self.torch_dtype = hyperparameters.get("torch_dtype", "auto" if self.min_num_gpus == 0 else "bfloat16")
+        self.torch_dtype = hyperparameters.get("torch_dtype", self.default_torch_dtype)
 
         self.data_loader_num_workers = hyperparameters.get("data_loader_num_workers", 0)
         self.optimization_strategy: Optional[Literal["onnx", "openvino"]] = hyperparameters.get(
@@ -200,8 +217,32 @@ class ChronosModel(AbstractTimeSeriesModel):
         return torch.cuda.is_available()
 
     @property
-    def min_num_gpus(self):
-        return MODEL_CONFIGS.get(self.model_path, {}).get("num_gpus", 0)
+    def ag_default_config(self) -> Dict[str, Any]:
+        """The default configuration of the model used by AutoGluon if the model is one of those
+        defined in MODEL_CONFIGS. For now, these are ``amazon/chronos-t5-*`` family of models.
+        """
+        return MODEL_CONFIGS.get(self.model_path, {})
+
+    @property
+    def min_num_gpus(self) -> int:
+        """Minimum number of GPUs required for the model. For models not defined in AutoGluon,
+        this value defaults to 0.
+        """
+        return self.ag_default_config.get("num_gpus", 0)
+
+    @property
+    def default_batch_size(self) -> int:
+        """Default batch size used for the model. For models not defined in AutoGluon, this value
+        defaults to 8.
+        """
+        return self.ag_default_config.get("default_batch_size", 8)
+
+    @property
+    def default_torch_dtype(self) -> Any:
+        """Default torch data type used for the model. For models not defined in AutoGluon, this value
+        defaults to "auto".
+        """
+        return self.ag_default_config.get("default_torch_dtype", "auto")
 
     def get_minimum_resources(self, is_gpu_available: bool = False) -> Dict[str, Union[int, float]]:
         minimum_resources = {"num_cpus": 1}
@@ -211,7 +252,7 @@ class ChronosModel(AbstractTimeSeriesModel):
         return minimum_resources
 
     def load_model_pipeline(self, context_length: Optional[int] = None):
-        from .chronos import OptimizedChronosPipeline
+        from .pipeline import OptimizedChronosPipeline
 
         gpu_available = self._is_gpu_available()
 
@@ -234,6 +275,10 @@ class ChronosModel(AbstractTimeSeriesModel):
 
         self.model_pipeline = pipeline
 
+    def persist(self) -> "ChronosModel":
+        self.load_model_pipeline(context_length=self.context_length or self.maximum_context_length)
+        return self
+
     def _fit(
         self,
         train_data: TimeSeriesDataFrame,
@@ -283,8 +328,9 @@ class ChronosModel(AbstractTimeSeriesModel):
         with warning_filter(all_warnings=True):
             import torch
 
-            # load model pipeline to device memory
-            self.load_model_pipeline(context_length=context_length)
+            if self.model_pipeline is None:
+                # load model pipeline to device memory
+                self.load_model_pipeline(context_length=context_length)
 
             self.model_pipeline.model.eval()
             with torch.inference_mode():
@@ -317,3 +363,6 @@ class ChronosModel(AbstractTimeSeriesModel):
         )
 
         return TimeSeriesDataFrame(df)
+
+    def _more_tags(self) -> Dict:
+        return {"allow_nan": True}

autogluon.timeseries-1.0.1b20240327/src/autogluon/timeseries/models/chronos/chronos.py → autogluon.timeseries-1.0.1b20240403/src/autogluon/timeseries/models/chronos/pipeline.py

@@ -2,7 +2,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
 # Original Source: https://github.com/amazon-science/chronos-forecasting
-# Author: Lorenzo Stella <stellalo@amazon.com>
+# Authors: Lorenzo Stella <stellalo@amazon.com>, Abdul Fatir Ansari <ansarnd@amazon.com>
 
 import logging
 import warnings
@@ -18,6 +18,9 @@ from autogluon.timeseries.utils.warning_filters import set_loggers_level
 logger = logging.getLogger(__name__)
 
 
+__all__ = ["ChronosConfig", "ChronosPipeline", "OptimizedChronosPipeline"]
+
+
 @dataclass
 class ChronosConfig:
     """
@@ -81,14 +84,14 @@ class ChronosTokenizer:
             A boolean tensor, same shape as ``token_ids``, indicating
             which input observations are not ``torch.nan`` (i.e. not
             missing nor padding).
-        decoding_context
+        tokenizer_state
             An object that will be passed to ``output_transform``.
             Contains the relevant context to decode output samples into
             real values, such as location and scale parameters.
         """
         raise NotImplementedError()
 
-    def output_transform(self, samples: torch.Tensor, decoding_context: Any) -> torch.Tensor:
+    def output_transform(self, samples: torch.Tensor, tokenizer_state: Any) -> torch.Tensor:
         """
         Turn a batch of sample token IDs into real values.
 
@@ -97,7 +100,7 @@ class ChronosTokenizer:
         samples
             A tensor of integers, shaped (batch_size, num_samples, time_length),
             containing token IDs of sample trajectories.
-        decoding_context
+        tokenizer_state
             An object returned by ``input_transform`` containing
             relevant context to decode samples, such as location and scale.
             The nature of this depends on the specific tokenizer.
@@ -132,13 +135,6 @@ class MeanScaleUniformBins(ChronosTokenizer):
 
         if length > self.config.context_length:
             context = context[..., -self.config.context_length :]
-        elif length < self.config.context_length:
-            padding_size = (
-                *context.shape[:-1],
-                self.config.context_length - length,
-            )
-            padding = torch.full(size=padding_size, fill_value=torch.nan)
-            context = torch.concat((padding, context), dim=-1)
 
         attention_mask = ~torch.isnan(context)
         scale = torch.nansum(torch.abs(context) * attention_mask, dim=-1) / torch.nansum(attention_mask, dim=-1)
@@ -191,7 +187,36 @@ class ChronosPretrainedModel(nn.Module):
         super().__init__()
         self.config = config
         self.model = model
-        self.device = model.device
+
+    @property
+    def device(self):
+        return self.model.device
+
+    def encode(
+        self,
+        input_ids: torch.Tensor,
+        attention_mask: torch.Tensor,
+    ):
+        """
+        Extract the encoder embedding for the given token sequences.
+
+        Parameters
+        ----------
+        input_ids
+            Tensor of indices of input sequence tokens in the vocabulary
+            with shape (batch_size, sequence_length).
+        attention_mask
+            A mask tensor of the same shape as input_ids to avoid attending
+            on padding or missing tokens.
+
+        Returns
+        -------
+        embedding
+            A tensor of encoder embeddings with shape
+            (batch_size, sequence_length, d_model).
+        """
+        assert self.config.model_type == "seq2seq", "Encoder embeddings are only supported for encoder-decoder models"
+        return self.model.encoder(input_ids=input_ids, attention_mask=attention_mask).last_hidden_state
 
     def forward(
         self,
@@ -288,6 +313,48 @@ class ChronosPipeline:
         self.tokenizer = tokenizer
         self.model = model
 
+    def _prepare_and_validate_context(self, context: Union[torch.Tensor, List[torch.Tensor]]):
+        if isinstance(context, list):
+            context = left_pad_and_stack_1D(context)
+        assert isinstance(context, torch.Tensor)
+        if context.ndim == 1:
+            context = context.unsqueeze(0)
+        assert context.ndim == 2
+
+        return context
+
+    @torch.no_grad()
+    def embed(self, context: Union[torch.Tensor, List[torch.Tensor]]) -> Tuple[torch.Tensor, Any]:
+        """
+        Get encoder embeddings for the given time series.
+
+        Parameters
+        ----------
+        context
+            Input series. This is either a 1D tensor, or a list
+            of 1D tensors, or a 2D tensor whose first dimension
+            is batch. In the latter case, use left-padding with
+            ``torch.nan`` to align series of different lengths.
+
+        Returns
+        -------
+        embeddings, tokenizer_state
+            A tuple of two tensors: the encoder embeddings and the tokenizer_state,
+            e.g., the scale of the time series in the case of mean scaling.
+            The encoder embeddings are shaped (batch_size, context_length, d_model)
+            or (batch_size, context_length + 1, d_model), where context_length
+            is the size of the context along the time axis if a 2D tensor was provided
+            or the length of the longest time series, if a list of 1D tensors was
+            provided, and the extra 1 is for EOS.
+        """
+        context = self._prepare_and_validate_context(context=context)
+        token_ids, attention_mask, tokenizer_state = self.tokenizer.input_transform(context)
+        embeddings = self.model.encode(
+            input_ids=token_ids.to(self.model.device),
+            attention_mask=attention_mask.to(self.model.device),
+        ).cpu()
+        return embeddings, tokenizer_state
+
     def predict(
         self,
         context: Union[torch.Tensor, List[torch.Tensor]],
@@ -335,13 +402,7 @@ class ChronosPipeline:
             Tensor of sample forecasts, of shape
             (batch_size, num_samples, prediction_length).
         """
-        if isinstance(context, list):
-            context = left_pad_and_stack_1D(context)
-        assert isinstance(context, torch.Tensor)
-        if context.ndim == 1:
-            context = context.unsqueeze(0)
-        assert context.ndim == 2
-
+        context = self._prepare_and_validate_context(context=context)
         if prediction_length is None:
             prediction_length = self.model.config.prediction_length
 

{autogluon.timeseries-1.0.1b20240327 → autogluon.timeseries-1.0.1b20240403}/src/autogluon/timeseries/models/gluonts/abstract_gluonts.py

@@ -328,8 +328,6 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
 
             if self.num_feat_static_real > 0:
                 feat_static_real = time_series_df.static_features[self.metadata.static_features_real]
-                if feat_static_real.isna().values.any():
-                    feat_static_real = feat_static_real.fillna(feat_static_real.mean())
             else:
                 feat_static_real = None
 
@@ -548,3 +546,6 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
 
         forecast_df.index = forecast_index
         return TimeSeriesDataFrame(forecast_df)
+
+    def _more_tags(self) -> dict:
+        return {"allow_nan": True, "can_use_val_data": True}

{autogluon.timeseries-1.0.1b20240327 → autogluon.timeseries-1.0.1b20240403}/src/autogluon/timeseries/models/local/abstract_local_model.py

@@ -1,7 +1,7 @@
 import logging
 import time
 from multiprocessing import TimeoutError, cpu_count
-from typing import Any, Dict, List, Optional, Tuple, Union
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 
 import numpy as np
 import pandas as pd
@@ -85,6 +85,12 @@ class AbstractLocalModel(AbstractTimeSeriesModel):
         self._local_model_args: Dict[str, Any] = None
         self._seasonal_period: Optional[int] = None
         self.time_limit: Optional[float] = None
+        self._dummy_forecast: Optional[pd.DataFrame] = None
+
+    def preprocess(self, data: TimeSeriesDataFrame, is_train: bool = False, **kwargs) -> Any:
+        if not self._get_tags()["allow_nan"]:
+            data = data.fill_missing_values()
+        return data
 
     def _fit(self, train_data: TimeSeriesDataFrame, time_limit: Optional[int] = None, **kwargs):
         self._check_fit_params()
@@ -115,8 +121,16 @@ class AbstractLocalModel(AbstractTimeSeriesModel):
 
         self._local_model_args = self._update_local_model_args(local_model_args=local_model_args)
         self.time_limit = time_limit
+
+        self._dummy_forecast = self._get_dummy_forecast(train_data)
         return self
 
+    def _get_dummy_forecast(self, train_data: TimeSeriesDataFrame) -> pd.DataFrame:
+        agg_functions = ["mean"] + [get_quantile_function(q) for q in self.quantile_levels]
+        stats_marginal = train_data[self.target].agg(agg_functions)
+        stats_repeated = np.tile(stats_marginal.values, [self.prediction_length, 1])
+        return pd.DataFrame(stats_repeated, columns=stats_marginal.index)
+
     def _update_local_model_args(self, local_model_args: Dict[str, Any]) -> Dict[str, Any]:
         return local_model_args
 
@@ -164,25 +178,30 @@ class AbstractLocalModel(AbstractTimeSeriesModel):
     def _predict_wrapper(self, time_series: pd.Series, end_time: Optional[float] = None) -> Tuple[pd.DataFrame, bool]:
         if end_time is not None and time.time() >= end_time:
             raise TimeLimitExceeded
-        try:
-            result = self._predict_with_local_model(
-                time_series=time_series,
-                local_model_args=self._local_model_args.copy(),
-            )
-            if not np.isfinite(result.values).all():
-                raise RuntimeError("Forecast contains NaN or Inf values.")
-            model_failed = False
-        except Exception:
-            if self.use_fallback_model:
-                result = seasonal_naive_forecast(
-                    target=time_series.values.ravel(),
-                    prediction_length=self.prediction_length,
-                    quantile_levels=self.quantile_levels,
-                    seasonal_period=self._seasonal_period,
+
+        if time_series.isna().all():
+            result = self._dummy_forecast.copy()
+            model_failed = True
+        else:
+            try:
+                result = self._predict_with_local_model(
+                    time_series=time_series,
+                    local_model_args=self._local_model_args.copy(),
                 )
-                model_failed = True
-            else:
-                raise
+                if not np.isfinite(result.values).all():
+                    raise RuntimeError("Forecast contains NaN or Inf values.")
+                model_failed = False
+            except Exception:
+                if self.use_fallback_model:
+                    result = seasonal_naive_forecast(
+                        target=time_series.values.ravel(),
+                        prediction_length=self.prediction_length,
+                        quantile_levels=self.quantile_levels,
+                        seasonal_period=self._seasonal_period,
+                    )
+                    model_failed = True
+                else:
+                    raise
         return result, model_failed
 
     def _predict_with_local_model(
@@ -197,25 +216,51 @@ def seasonal_naive_forecast(
     target: np.ndarray, prediction_length: int, quantile_levels: List[float], seasonal_period: int
 ) -> pd.DataFrame:
     """Generate seasonal naive forecast, predicting the last observed value from the same period."""
+
+    def numpy_ffill(arr: np.ndarray) -> np.ndarray:
+        """Fast implementation of forward fill in numpy."""
+        idx = np.arange(len(arr))
+        mask = np.isnan(arr)
+        idx[mask] = 0
+        return arr[np.maximum.accumulate(idx)]
+
     forecast = {}
+    # Convert to float64 since std computation can be unstable in float32
+    target = target.astype(np.float64)
     # At least seasonal_period + 2 values are required to compute sigma for seasonal naive
     if len(target) > seasonal_period + 1 and seasonal_period > 1:
+        if np.isnan(target[-(seasonal_period + 2) :]).any():
+            target = numpy_ffill(target)
+
         indices = [len(target) - seasonal_period + k % seasonal_period for k in range(prediction_length)]
         forecast["mean"] = target[indices]
         residuals = target[seasonal_period:] - target[:-seasonal_period]
 
-        sigma = np.sqrt(np.mean(np.square(residuals)))
+        sigma = np.sqrt(np.nanmean(np.square(residuals)))
         num_full_seasons = np.arange(1, prediction_length + 1) // seasonal_period
         sigma_per_timestep = sigma * np.sqrt(num_full_seasons + 1)
     else:
         # Fall back to naive forecast
-        forecast["mean"] = np.full(shape=[prediction_length], fill_value=target[-1])
+        last_observed_value = target[np.isfinite(target)][-1]
+        forecast["mean"] = np.full(shape=[prediction_length], fill_value=last_observed_value)
         residuals = target[1:] - target[:-1]
 
-        sigma = np.sqrt(np.mean(np.square(residuals)))
+        sigma = np.sqrt(np.nanmean(np.square(residuals)))
+        if np.isnan(sigma):  # happens if there are no two consecutive non-nan observations
+            sigma = 0.0
         sigma_per_timestep = sigma * np.sqrt(np.arange(1, prediction_length + 1))
 
     for q in quantile_levels:
         forecast[str(q)] = forecast["mean"] + norm.ppf(q) * sigma_per_timestep
 
     return pd.DataFrame(forecast)
+
+
+def get_quantile_function(q: float) -> Callable:
+    """Returns a function with name "q" that computes the q'th quantile of a pandas.Series."""
+
+    def quantile_fn(x: pd.Series) -> pd.Series:
+        return x.quantile(q)
+
+    quantile_fn.__name__ = str(q)
+    return quantile_fn