Pros of Prophet

• Widely adopted and battle-tested in production environments
• Extensive documentation and community support
• Handles holidays and seasonal effects out-of-the-box

Cons of Prophet

• Less flexible for custom model architectures
• Can be slower for large-scale forecasting tasks
• Limited support for external regressors

Code Comparison

Prophet:

from fbprophet import Prophet

m = Prophet()
m.fit(df)
future = m.make_future_dataframe(periods=365)
forecast = m.predict(future)

NeuralForecast:

from neuralforecast import NeuralForecast
from neuralforecast.models import NBEATS

model = NeuralForecast(models=[NBEATS(h=24, input_size=24)])
model.fit(df)
forecast = model.predict()

NeuralForecast offers a more flexible approach to time series forecasting, leveraging neural network architectures. It provides better performance for large-scale forecasting tasks and allows for easier customization of model architectures. However, Prophet remains a solid choice for simpler forecasting tasks and benefits from its extensive documentation and community support.

Pros of sktime

• Broader scope covering various time series tasks beyond forecasting
• More extensive documentation and tutorials
• Larger community and ecosystem of extensions

Cons of sktime

• Steeper learning curve due to more complex architecture
• Slower performance for some tasks compared to specialized libraries

Code Comparison

sktime:

from sktime.forecasting.naive import NaiveForecaster
from sktime.datasets import load_airline

y = load_airline()
forecaster = NaiveForecaster(strategy="mean")
forecaster.fit(y)
y_pred = forecaster.predict(fh=[1,2,3])

neuralforecast:

from neuralforecast import NeuralForecast
from neuralforecast.models import NBEATS

model = NeuralForecast(models=[NBEATS(input_size=30, h=5)])
model.fit(df)
forecast = model.predict()

The code examples highlight the different approaches:

• sktime focuses on a unified interface for various forecasting methods
• neuralforecast specializes in neural network-based forecasting models

Both libraries offer powerful forecasting capabilities, but sktime provides a more comprehensive toolkit for time series analysis, while neuralforecast focuses on state-of-the-art neural network models for forecasting.

Pros of darts

• Broader range of models and algorithms, including classical statistical methods
• More extensive documentation and tutorials
• Supports both univariate and multivariate time series forecasting

Cons of darts

• Generally slower performance, especially for large datasets
• Less focus on deep learning models compared to neuralforecast
• May require more manual feature engineering for some models

Code Comparison

darts:

from darts import TimeSeries
from darts.models import Prophet

series = TimeSeries.from_dataframe(df, 'date', 'value')
model = Prophet()
model.fit(series)
forecast = model.predict(n=30)

neuralforecast:

from neuralforecast import NeuralForecast
from neuralforecast.models import NBEATS

model = NeuralForecast(models=[NBEATS(input_size=30, h=7)])
model.fit(df)
forecast = model.predict()

Both libraries offer intuitive APIs for time series forecasting, but neuralforecast focuses more on deep learning models and provides a more streamlined approach for neural network-based forecasting. darts offers a wider variety of traditional and machine learning models, making it more versatile for different types of time series problems.

Pros of pytorch-forecasting

• More extensive documentation and examples
• Wider range of models and architectures available
• Better integration with PyTorch ecosystem

Cons of pytorch-forecasting

• Steeper learning curve for beginners
• Less focus on traditional statistical methods
• May be overkill for simpler forecasting tasks

Code Comparison

neuralforecast:

from neuralforecast import NeuralForecast
from neuralforecast.models import NBEATS

model = NeuralForecast(models=[NBEATS(input_size=7, h=1, max_epochs=50)])
model.fit(df)
forecast = model.predict()

pytorch-forecasting:

from pytorch_forecasting import TemporalFusionTransformer, TimeSeriesDataSet

training = TimeSeriesDataSet(
    data,
    time_idx="timestamp",
    target="target",
    group_ids=["id"],
    static_categoricals=["category"],
    time_varying_known_reals=["price"],
    time_varying_unknown_reals=["weather"],
)
model = TemporalFusionTransformer.from_dataset(training)
model.fit(train_dataloader, val_dataloader)
predictions = model.predict(test_dataloader)

The code comparison shows that pytorch-forecasting requires more setup and configuration, but offers greater flexibility in defining the dataset and model architecture. neuralforecast provides a simpler API for quick implementation of common forecasting tasks.

Pros of Orbit

• Offers a wider range of probabilistic time series models, including Bayesian structural time series
• Provides built-in diagnostics and model selection tools
• Supports Bayesian inference for parameter estimation and uncertainty quantification

Cons of Orbit

• Less focus on deep learning models compared to NeuralForecast
• May have a steeper learning curve for users unfamiliar with Bayesian methods
• Potentially slower training and inference times for large-scale forecasting tasks

Code Comparison

NeuralForecast:

from neuralforecast import NeuralForecast
from neuralforecast.models import NBEATS

model = NeuralForecast(models=[NBEATS(h=24, input_size=72)])
model.fit(df)
forecast = model.predict()

Orbit:

from orbit.models import DLT

model = DLT(response_col='y', date_col='ds', seasonality=52)
model.fit(df)
prediction = model.predict(df)

Both libraries offer concise APIs for model creation, fitting, and prediction. NeuralForecast focuses on neural network-based models, while Orbit provides a broader range of traditional and Bayesian time series models. The choice between them depends on the specific forecasting needs and the user's familiarity with different modeling approaches.