durbyn

Python wrapper for Durbyn.jl time series forecasting

---

durbyn brings the power of the Julia Durbyn.jl time series forecasting package to Python. It provides a familiar, scikit-learn-style API with .fit() / .forecast() methods, numpy arrays, and optional pandas integration — while delegating all computation to Julia for performance.

Durbyn — Kurdish for "binoculars" (Dur, far + Byn, to see), embodies foresight through science.

Features

• 21 forecasting models — Exponential Smoothing (SES, Holt, Holt-Winters, ETS), ARIMA, AutoARIMA, BATS, TBATS, Theta, Diffusion, ARAR, ARARMA, Croston variants, and Naive baselines
• Scikit-learn-style API — .fit(y, m=12) → .forecast(h=12, level=[80, 95])
• Prediction intervals — configurable confidence levels
• Model comparison — compare() evaluates multiple models on train/test splits
• Panel data — PanelForecaster fits per group in a DataFrame
• Pure Python metrics — ME, RMSE, MAE, MPE, MAPE, MASE, ACF1
• Pandas integration — .to_dataframe() on all results
• Matplotlib plotting — .plot() with history and prediction intervals

Installation

pip install durbyn

Optional dependencies:

pip install durbyn[all]     # pandas + matplotlib
pip install durbyn[pandas]  # DataFrame support
pip install durbyn[plot]    # plotting

Note: Julia must be installed on your system. The juliacall package manages the Julia runtime automatically.

Quick Start

from durbyn import AutoARIMA

# Fit an automatic ARIMA model
model = AutoARIMA().fit(y, m=12)

# Generate forecasts with prediction intervals
fc = model.forecast(h=12, level=[80, 95])

print(fc.mean)            # numpy array of point forecasts
print(fc.to_dataframe())  # pandas DataFrame
fc.plot()                 # matplotlib figure

Models

Category	Classes
Exponential Smoothing	SES, Holt, HoltWinters, ETS, Croston
ARIMA	ARIMA, AutoARIMA
Naive	Naive, SeasonalNaive, RandomWalk, MeanForecast
BATS	BATS, TBATS
Theta	Theta, AutoTheta
Diffusion	Diffusion
ARAR	ARAR, ARARMA, AutoARARMA
Intermittent Demand	CrostonClassic, CrostonSBA, CrostonSBJ

All models follow the same interface:

model = SomeModel(**params)
model.fit(y, m=seasonal_period)
fc = model.forecast(h=steps_ahead, level=[80, 95])

fc.mean              # point forecasts
fc.lower[80]         # lower bound at 80%
fc.upper[95]         # upper bound at 95%
model.fitted_values  # in-sample fitted values
model.residuals      # in-sample residuals
model.summary        # model summary string

Model Comparison

from durbyn import SES, Holt, HoltWinters, AutoARIMA, compare

result = compare(
    models={
        "SES": SES(),
        "Holt": Holt(damped=True),
        "HW": HoltWinters(seasonal="multiplicative"),
        "AutoARIMA": AutoARIMA(),
    },
    y_train=y_train,
    y_test=y_test,
    m=12,
)

print(result.to_dataframe())  # accuracy metrics per model
result.plot()                  # overlay all forecasts

Panel Data

import pandas as pd
from durbyn import AutoARIMA, PanelForecaster

pf = PanelForecaster(model=AutoARIMA(), groupby="store")
pf.fit(data, y_col="sales", m=12, date_col="date")
result = pf.forecast(h=12, level=[80, 95])
print(result.to_dataframe())

Documentation

Full documentation: taf-society.github.io/durbyn

About TAFS

TAFS (Time Series Analysis and Forecasting Society) is a non-profit association in Vienna, Austria, connecting academics, experts, practitioners, and students focused on time-series, forecasting, and decision science. Learn more at taf-society.org.

License

MIT License. See LICENSE for details.