API Reference¶

Complete reference for all public classes and functions.

Models¶

All models follow the same interface:

model = ModelClass(**params)       # configure
model.fit(y, m=1, X=None)         # fit to data → returns self
fc = model.forecast(h=10, level=None)  # forecast → ForecastResult

Properties (after fitting)¶

Property	Type	Description
`fitted_values`	`np.ndarray`	In-sample fitted values
`residuals`	`np.ndarray`	In-sample residuals
`summary`	`str`	Model summary string
`coefficients`	`dict`	Model coefficients

Methods¶

Method	Returns	Description
`.fit(y, m=1, X=None)`	`self`	Fit model to data
`.forecast(h=10, level=None)`	`ForecastResult`	Generate forecasts
`.predict(h=10, **kw)`	`ForecastResult`	Alias for `.forecast()`
`.get_params()`	`dict`	Return model parameters

Exponential Smoothing¶

`SES(initial="optimal", alpha=None, lambda_=None, biasadj=False)`¶

Simple Exponential Smoothing. User Guide

`Holt(damped=False, initial="optimal", exponential=False, alpha=None, beta=None, phi=None, lambda_=None, biasadj=False)`¶

Holt's linear trend method. User Guide

`HoltWinters(seasonal="additive", damped=False, initial="optimal", exponential=False, alpha=None, beta=None, gamma=None, phi=None, lambda_=None, biasadj=False)`¶

Holt-Winters triple exponential smoothing. User Guide

`ETS(model="ZZZ", damped=None, alpha=None, beta=None, gamma=None, phi=None, additive_only=False, lambda_=None, biasadj=False, opt_crit="lik", ic="aicc", bounds="both", restrict=True, allow_multiplicative_trend=False)`¶

Error-Trend-Seasonal state space model. User Guide

`Croston()`¶

Simple Croston method for intermittent demand. User Guide

ARIMA¶

`ARIMA(order=(0,0,0), seasonal_order=None, include_mean=True, method="CSS-ML", transform_pars=True, lambda_=None, biasadj=None)`¶

ARIMA/SARIMA model. User Guide

`AutoARIMA(d=None, D=None, max_p=5, max_q=5, max_P=2, max_Q=2, max_order=5, max_d=2, max_D=1, stationary=False, seasonal=True, stepwise=True, trace=False, approximation=True, ic="aicc", test="adf", seasonal_test="ocsb", allowmean=True, allowdrift=True, lambda_=None, biasadj=None)`¶

Automatic ARIMA selection. User Guide

Naive Methods¶

`Naive(lambda_=None, biasadj=False)`¶

Naive forecast (last value). User Guide

`SeasonalNaive(lambda_=None, biasadj=False)`¶

Seasonal naive forecast. User Guide

`RandomWalk(drift=False, lambda_=None, biasadj=False)`¶

Random walk with optional drift. User Guide

`MeanForecast(lambda_=None, biasadj=False)`¶

Mean forecast. User Guide

BATS / TBATS¶

`BATS(use_box_cox=None, use_trend=None, use_damped_trend=None, use_arma_errors=True, bc_lower=0.0, bc_upper=1.0, biasadj=False)`¶

BATS model for complex seasonality. User Guide

`TBATS(use_box_cox=None, use_trend=None, use_damped_trend=None, use_arma_errors=True, bc_lower=0.0, bc_upper=1.0, biasadj=False)`¶

Trigonometric BATS for non-integer periods. User Guide

Theta¶

`Theta(model_type="OTM", alpha=None, theta_param=None, initial_level=None, nmse=3)`¶

Theta method. User Guide

`AutoTheta(model=None, decomposition_type="multiplicative", alpha=None, theta_param=None, initial_level=None, nmse=3)`¶

Automatic Theta selection. User Guide

Diffusion¶

`Diffusion(model_type="Bass", cleanlead=True, loss=2, cumulative=True)`¶

Innovation diffusion model. User Guide

ARAR / ARARMA¶

`ARAR(max_ar_depth=None, max_lag=None)`¶

ARAR model. User Guide

`ARARMA(max_ar_depth=26, max_lag=40, p=4, q=1)`¶

ARARMA model. User Guide

`AutoARARMA(min_p=0, max_p=4, min_q=0, max_q=2, max_ar_depth=26, max_lag=40)`¶

Automatic ARARMA selection. User Guide

Intermittent Demand¶

`CrostonClassic(init_strategy="mean", number_of_params=2, cost_metric="mar", optimize_init=True, rm_missing=False)`¶

Croston's classic method. User Guide

`CrostonSBA(...)`¶

Syntetos-Boylan Approximation. Same parameters as CrostonClassic. User Guide

`CrostonSBJ(...)`¶

Shale-Boylan-Johnston variant. Same parameters as CrostonClassic. User Guide

ForecastResult¶

@dataclass
class ForecastResult:
    mean: np.ndarray              # point forecasts
    lower: dict[int, np.ndarray]  # {80: [...], 95: [...]}
    upper: dict[int, np.ndarray]  # {80: [...], 95: [...]}
    level: list[int]              # [80, 95]
    fitted_values: np.ndarray     # in-sample fitted
    residuals: np.ndarray         # in-sample residuals
    method: str                   # model description
    x: np.ndarray                 # original data

Methods¶

Method	Returns	Description
`.to_dataframe()`	`pd.DataFrame`	Columns: step, mean, lower_80, upper_80, ...
`.plot(ax=None, show_history=True)`	matplotlib `Axes`	Plot forecast with intervals

compare()¶

compare(
    models: dict[str, BaseForecaster],
    y_train: array-like,
    y_test: array-like,
    m: int = 1,
    h: int | None = None,
    level: list[int] | None = None,
) -> ComparisonResult

User Guide

ComparisonResult¶

Attribute	Type	Description
`model_names`	`list[str]`	Model names
`forecasts`	`dict[str, ForecastResult]`	Forecasts per model
`metrics`	`dict[str, dict[str, float]]`	Metrics per model

Method	Returns	Description
`.to_dataframe()`	`pd.DataFrame`	One row per model, metric columns
`.plot(ax=None)`	matplotlib `Axes`	Overlay all forecasts

PanelForecaster¶

PanelForecaster(model: BaseForecaster, groupby: str | list[str])

Method	Returns	Description
`.fit(data, y_col="y", m=1, date_col=None)`	`self`	Fit per group
`.forecast(h=10, level=None)`	`PanelForecastResult`	Forecast all groups

User Guide

PanelForecastResult¶

Attribute	Type	Description
`groups`	`dict[Any, ForecastResult]`	Results per group

Method	Returns	Description
`.to_dataframe()`	`pd.DataFrame`	Long-format with group column

Metrics¶

from durbyn.metrics import me, rmse, mae, mpe, mape, mase, acf1, accuracy

Function	Signature	Description
`me(actual, predicted)`	`→ float`	Mean Error
`rmse(actual, predicted)`	`→ float`	Root Mean Squared Error
`mae(actual, predicted)`	`→ float`	Mean Absolute Error
`mpe(actual, predicted)`	`→ float`	Mean Percentage Error (%)
`mape(actual, predicted)`	`→ float`	Mean Absolute Percentage Error (%)
`mase(actual, predicted, training_data)`	`→ float`	Mean Absolute Scaled Error
`acf1(actual, predicted)`	`→ float`	Lag-1 autocorrelation of errors
`accuracy(actual, predicted, training_data=None)`	`→ dict`	All metrics combined

User Guide

Exceptions¶

from durbyn.exceptions import DurbynError, NotFittedError, JuliaNotAvailableError, DurbynNotInstalledError

Exception	Parent	Description
`DurbynError`	`Exception`	Base exception
`JuliaNotAvailableError`	`DurbynError, ImportError`	Julia runtime not found
`DurbynNotInstalledError`	`DurbynError, ImportError`	Durbyn.jl not installed
`NotFittedError`	`DurbynError, RuntimeError`	Model not fitted yet

API Reference¶

Models¶

Properties (after fitting)¶

Methods¶

Exponential Smoothing¶

SES(initial="optimal", alpha=None, lambda_=None, biasadj=False)¶

Holt(damped=False, initial="optimal", exponential=False, alpha=None, beta=None, phi=None, lambda_=None, biasadj=False)¶

HoltWinters(seasonal="additive", damped=False, initial="optimal", exponential=False, alpha=None, beta=None, gamma=None, phi=None, lambda_=None, biasadj=False)¶

ETS(model="ZZZ", damped=None, alpha=None, beta=None, gamma=None, phi=None, additive_only=False, lambda_=None, biasadj=False, opt_crit="lik", ic="aicc", bounds="both", restrict=True, allow_multiplicative_trend=False)¶

Croston()¶

ARIMA¶

ARIMA(order=(0,0,0), seasonal_order=None, include_mean=True, method="CSS-ML", transform_pars=True, lambda_=None, biasadj=None)¶

AutoARIMA(d=None, D=None, max_p=5, max_q=5, max_P=2, max_Q=2, max_order=5, max_d=2, max_D=1, stationary=False, seasonal=True, stepwise=True, trace=False, approximation=True, ic="aicc", test="adf", seasonal_test="ocsb", allowmean=True, allowdrift=True, lambda_=None, biasadj=None)¶

Naive Methods¶

Naive(lambda_=None, biasadj=False)¶

SeasonalNaive(lambda_=None, biasadj=False)¶

RandomWalk(drift=False, lambda_=None, biasadj=False)¶

MeanForecast(lambda_=None, biasadj=False)¶

BATS / TBATS¶

BATS(use_box_cox=None, use_trend=None, use_damped_trend=None, use_arma_errors=True, bc_lower=0.0, bc_upper=1.0, biasadj=False)¶

TBATS(use_box_cox=None, use_trend=None, use_damped_trend=None, use_arma_errors=True, bc_lower=0.0, bc_upper=1.0, biasadj=False)¶

Theta¶

Theta(model_type="OTM", alpha=None, theta_param=None, initial_level=None, nmse=3)¶

AutoTheta(model=None, decomposition_type="multiplicative", alpha=None, theta_param=None, initial_level=None, nmse=3)¶

Diffusion¶

Diffusion(model_type="Bass", cleanlead=True, loss=2, cumulative=True)¶

ARAR / ARARMA¶

ARAR(max_ar_depth=None, max_lag=None)¶

ARARMA(max_ar_depth=26, max_lag=40, p=4, q=1)¶

AutoARARMA(min_p=0, max_p=4, min_q=0, max_q=2, max_ar_depth=26, max_lag=40)¶

Intermittent Demand¶

CrostonClassic(init_strategy="mean", number_of_params=2, cost_metric="mar", optimize_init=True, rm_missing=False)¶

CrostonSBA(...)¶

CrostonSBJ(...)¶

ForecastResult¶

Methods¶

compare()¶

ComparisonResult¶

PanelForecaster¶

PanelForecastResult¶

Metrics¶

Exceptions¶

`SES(initial="optimal", alpha=None, lambda_=None, biasadj=False)`¶

`Holt(damped=False, initial="optimal", exponential=False, alpha=None, beta=None, phi=None, lambda_=None, biasadj=False)`¶

`HoltWinters(seasonal="additive", damped=False, initial="optimal", exponential=False, alpha=None, beta=None, gamma=None, phi=None, lambda_=None, biasadj=False)`¶

`ETS(model="ZZZ", damped=None, alpha=None, beta=None, gamma=None, phi=None, additive_only=False, lambda_=None, biasadj=False, opt_crit="lik", ic="aicc", bounds="both", restrict=True, allow_multiplicative_trend=False)`¶

`Croston()`¶

`ARIMA(order=(0,0,0), seasonal_order=None, include_mean=True, method="CSS-ML", transform_pars=True, lambda_=None, biasadj=None)`¶

`AutoARIMA(d=None, D=None, max_p=5, max_q=5, max_P=2, max_Q=2, max_order=5, max_d=2, max_D=1, stationary=False, seasonal=True, stepwise=True, trace=False, approximation=True, ic="aicc", test="adf", seasonal_test="ocsb", allowmean=True, allowdrift=True, lambda_=None, biasadj=None)`¶

`Naive(lambda_=None, biasadj=False)`¶

`SeasonalNaive(lambda_=None, biasadj=False)`¶

`RandomWalk(drift=False, lambda_=None, biasadj=False)`¶

`MeanForecast(lambda_=None, biasadj=False)`¶

`BATS(use_box_cox=None, use_trend=None, use_damped_trend=None, use_arma_errors=True, bc_lower=0.0, bc_upper=1.0, biasadj=False)`¶

`TBATS(use_box_cox=None, use_trend=None, use_damped_trend=None, use_arma_errors=True, bc_lower=0.0, bc_upper=1.0, biasadj=False)`¶

`Theta(model_type="OTM", alpha=None, theta_param=None, initial_level=None, nmse=3)`¶

`AutoTheta(model=None, decomposition_type="multiplicative", alpha=None, theta_param=None, initial_level=None, nmse=3)`¶

`Diffusion(model_type="Bass", cleanlead=True, loss=2, cumulative=True)`¶

`ARAR(max_ar_depth=None, max_lag=None)`¶

`ARARMA(max_ar_depth=26, max_lag=40, p=4, q=1)`¶

`AutoARARMA(min_p=0, max_p=4, min_q=0, max_q=2, max_ar_depth=26, max_lag=40)`¶

`CrostonClassic(init_strategy="mean", number_of_params=2, cost_metric="mar", optimize_init=True, rm_missing=False)`¶

`CrostonSBA(...)`¶

`CrostonSBJ(...)`¶