Model Catalog Overview

Source Files

• twiga/models/baseline/ - no-training baselines

• twiga/models/ml/ - tree-based and linear models

• twiga/models/nn/ - neural network models

• twiga/models/nn/rnn_model.py - RNNModel / RNNConfig

• twiga/models/nn/prob/rnn_parametric.py - RNN parametric distribution wrappers

• twiga/models/nn/prob/rnn_qr.py - RNNQR / RNNFPQR

• twiga/forecaster/registry.py

A single TwigaForecaster trains, tunes, evaluates and backtests any model in the catalog through an identical API. Switching from a gradient-boosted tree to a neural network means changing one config object, not rewriting training code. The entire workflow — feature engineering, training, hyperparameter optimisation, backtesting and calibration — is unified in one interface with no library boundaries to cross.

Twiga provides forecasting models across three domains: baseline (parameter-free reference models), ml (gradient-boosted trees, random forests, linear, and NGBoost), and nn (neural networks). All models share a common interface and are loaded dynamically through the model registry.

The table below lists the base model configurations. In addition, every neural architecture (MLPF, MLPGAM, MLPGAF, NHITS) has dedicated probabilistic config classes for each distribution family (Normal, Laplace, LogNormal, Gamma, Beta, StudentT, QR, FPQR, CRC) - see Neural Network Models for the full matrix.

Available Models

Baseline Domain

Model	Name	Domain	Type	Config Class	Description
Naive	naive	baseline	Point	NAIVEConfig	Per-window last value, training mean, or zero
Seasonal Naive	seasonal_naive	baseline	Point	SEASONALNAIVEConfig	Repeats value from m steps ago (daily/weekly)
Window Average	window_average	baseline	Point	WINDOWAVERAGEConfig	Local mean over the last w steps
Drift	drift	baseline	Point	DRIFTConfig	Linear extrapolation of per-window trend
Context Parrot	context_parrot	baseline	Point	CONTEXTPARROTConfig	Repeats the last observed context window as the forecast

ML Domain

Model	Name	Domain	Type	Config Class	Description
CatBoost	catboost	ml	Point	CATBOOSTConfig	Gradient boosting with ordered boosting
XGBoost	xgboost	ml	Point	XGBOOSTConfig	Gradient boosting with regularization
LightGBM	lightgbm	ml	Point	LIGHTGBMConfig	Histogram-based gradient boosting
Random Forest	randomforest	ml	Point	RANDOMFORESTConfig	Bagged decision trees, native multi-output
Linear Regression	lineareg	ml	Point	LINEAREGConfig	OLS linear model
Gaussian CatBoost	gausscatboost	ml	Probabilistic	GAUSSCATBOOSTConfig	CatBoost with Gaussian output (mean + sigma)
NGBoost Normal	ngboostnormal	ml	Probabilistic	NGBOOSTNORMALConfig	Natural gradient boosting, Gaussian distribution
NGBoost LogNormal	ngboostlognormal	ml	Probabilistic	NGBOOSTLOGNORMALConfig	Natural gradient boosting, LogNormal (positive targets)
NGBoost Exponential	ngboostexponential	ml	Probabilistic	NGBOOSTEXPONENTIALConfig	Natural gradient boosting, Exponential distribution
QR CatBoost	qrcatboost	ml	Quantile	QRCATBOOSTConfig	CatBoost with multi-quantile loss
QR XGBoost	qrxgboost	ml	Quantile	QRXGBOOSTConfig	XGBoost with quantile regression
QR LightGBM	qrlightgbm	ml	Quantile	QRLIGHTGBMConfig	LightGBM with quantile regression
QR Random Forest	qrrandomforest	ml	Quantile	QRRANDOMFORESTConfig	Single-forest quantile regression, any quantile at inference

NN Domain

Each architecture ships as a point base model plus a full suite of probabilistic variants. All share the nn domain and are loaded via the same registry.

Base Architectures

Model	Name	Config Class	Description
MLPF	mlpf	MLPFConfig	MLP forecaster with attention-based feature combination
MLPGAM	mlpgam	MLPGAMConfig	MLP with Generalized Additive Model structure
MLPGAF	mlpgaf	MLPGAFConfig	MLP with Gated Additive Features
N-HiTS	nhits	NHITSConfig	Neural Hierarchical Interpolation for Time Series
RNN	rnn	RNNConfig	GRU/LSTM recurrent network with bidirectional support

MLPF Probabilistic Variants

Name	Type	Config Class
mlpfnormal	Parametric — Normal	MLPFNormalConfig
mlpflaplace	Parametric — Laplace	MLPFLaplaceConfig
mlpflognormal	Parametric — LogNormal	MLPFLogNormalConfig
mlpfgamma	Parametric — Gamma	MLPFGammaConfig
mlpfbeta	Parametric — Beta	MLPFBetaConfig
mlpfqr	Quantile Regression	MLPFQRConfig
mlpffpqr	Full Parameterised QR	MLPFFPQRConfig
mlpfcrc	Conformal Residual Coverage	MLPFCRCConfig

MLPGAM Probabilistic Variants

Name	Type	Config Class
mlpgamnormal	Parametric — Normal	MLPGAMNormalConfig
mlpgamlaplace	Parametric — Laplace	MLPGAMLaplaceConfig
mlpgamlognormal	Parametric — LogNormal	MLPGAMLogNormalConfig
mlpgamgamma	Parametric — Gamma	MLPGAMGammaConfig
mlpgambeta	Parametric — Beta	MLPGAMBetaConfig
mlpgamstudentt	Parametric — Student-t	MLPGAMStudentTConfig
mlpgamqr	Quantile Regression	MLPGAMQRConfig
mlpgamfpqr	Full Parameterised QR	MLPGAMFPQRConfig
mlpgamcrc	Conformal Residual Coverage	MLPGAMCRCConfig

MLPGAF Probabilistic Variants

Name	Type	Config Class
mlpgafnormal	Parametric — Normal	MLPGAFNormalConfig
mlpgaflaplace	Parametric — Laplace	MLPGAFLaplaceConfig
mlpgaflognormal	Parametric — LogNormal	MLPGAFLogNormalConfig
mlpgafgamma	Parametric — Gamma	MLPGAFGammaConfig
mlpgafbeta	Parametric — Beta	MLPGAFBetaConfig
mlpgafstudentt	Parametric — Student-t	MLPGAFStudentTConfig
mlpgafqr	Quantile Regression	MLPGAFQRConfig
mlpgaffpqr	Full Parameterised QR	MLPGAFFPQRConfig
mlpgafcrc	Conformal Residual Coverage	MLPGAFCRCConfig

N-HiTS Probabilistic Variants

Name	Type	Config Class
nhitsnormal	Parametric — Normal	NHITSNormalConfig
nhitslaplace	Parametric — Laplace	NHITSLaplaceConfig
nhitslognormal	Parametric — LogNormal	NHITSLogNormalConfig
nhitsgamma	Parametric — Gamma	NHITSGammaConfig
nhitsbeta	Parametric — Beta	NHITSBetaConfig
nhitsstudentt	Parametric — Student-t	NHITSStudentTConfig
nhitsqr	Quantile Regression	NHITSQRConfig
nhitsfpqr	Full Parameterised QR	NHITSFPQRConfig
nhitscrc	Conformal Residual Coverage	NHITSCRCConfig

RNN Probabilistic Variants

Name	Type	Config Class
rnnnormal	Parametric — Normal	RNNNormalConfig
rnnlaplace	Parametric — Laplace	RNNLaplaceConfig
rnnlognormal	Parametric — LogNormal	RNNLogNormalConfig
rnngamma	Parametric — Gamma	RNNGammaConfig
rnnbeta	Parametric — Beta	RNNBetaConfig
rnnstudentt	Parametric — Student-t	RNNStudentTConfig
rnnqr	Quantile Regression	RNNQRConfig
rnnfpqr	Full Parameterised QR	RNNFPQRConfig

Model Class Hierarchy

        classDiagram
    class BaseRegressor {
        <<ML / Baseline>>
        +fit(X, y)
        +predict(x)
        +forecast(x)
    }

    class BaseQuantileRegressor {
        <<ML Quantile>>
        +quantiles: list
        +fit(X, y)
        +predict(X, sigma)
    }

    class BaseNGBoostRegressor {
        <<ML Probabilistic>>
        +fit(X, y)
        +predict(X)
    }

    class BaseNeuralForecast {
        <<NN Base>>
        +fit(X_train, y_train, X_val, y_val)
        +forecast(features)
        +update(trial)
    }

    class MLPFVariants["MLPF Variants (×8)"] {
        mlpfnormal · mlpflaplace · mlpflognormal
        mlpfgamma · mlpfbeta
        mlpfqr · mlpffpqr · mlpfcrc
    }

    class MLPGAMVariants["MLPGAM Variants (×9)"] {
        mlpgamnormal · mlpgamlaplace · mlpgamlognormal
        mlpgamgamma · mlpgambeta · mlpgamstudentt
        mlpgamqr · mlpgamfpqr · mlpgamcrc
    }

    class MLPGAFVariants["MLPGAF Variants (×9)"] {
        mlpgafnormal · mlpgaflaplace · mlpgaflognormal
        mlpgafgamma · mlpgafbeta · mlpgafstudentt
        mlpgafqr · mlpgaffpqr · mlpgafcrc
    }

    class NHITSVariants["N-HiTS Variants (×9)"] {
        nhitsnormal · nhitslaplace · nhitslognormal
        nhitsgamma · nhitsbeta · nhitsstudentt
        nhitsqr · nhitsfpqr · nhitscrc
    }

    class RNNVariants["RNN Variants (×8)"] {
        rnnnormal · rnnlaplace · rnnlognormal
        rnngamma · rnnbeta · rnnstudentt
        rnnqr · rnnfpqr
    }

    BaseRegressor <|-- NAIVEModel
    BaseRegressor <|-- SEASONALNAIVEModel
    BaseRegressor <|-- WINDOWAVERAGEModel
    BaseRegressor <|-- DRIFTModel
    BaseRegressor <|-- CONTEXTPARROTModel
    BaseRegressor <|-- CATBOOSTModel
    BaseRegressor <|-- XGBOOSTModel
    BaseRegressor <|-- LIGHTGBMModel
    BaseRegressor <|-- RANDOMFORESTModel
    BaseRegressor <|-- LINEAREGModel
    BaseRegressor <|-- GAUSSCATBOOSTModel
    BaseRegressor <|-- QRRANDOMFORESTModel
    BaseRegressor <|-- BaseQuantileRegressor
    BaseQuantileRegressor <|-- QRCATBOOSTModel
    BaseQuantileRegressor <|-- QRXGBOOSTModel
    BaseQuantileRegressor <|-- QRLIGHTGBMModel
    BaseNGBoostRegressor <|-- NGBOOSTNORMALModel
    BaseNGBoostRegressor <|-- NGBOOSTLOGNORMALModel
    BaseNGBoostRegressor <|-- NGBOOSTEXPONENTIALModel

    BaseNeuralForecast <|-- MLPFModel
    BaseNeuralForecast <|-- MLPGAMModel
    BaseNeuralForecast <|-- MLPGAFModel
    BaseNeuralForecast <|-- NHITSModel
    BaseNeuralForecast <|-- RNNModel
    MLPFModel <|-- MLPFVariants
    MLPGAMModel <|-- MLPGAMVariants
    MLPGAFModel <|-- MLPGAFVariants
    NHITSModel <|-- NHITSVariants
    RNNModel <|-- RNNVariants
    

Choosing a Model

Use Case	Recommended Models	Why
Establish performance floor	naive, seasonal_naive, drift, context_parrot	Zero-training reference; compute skill scores against ML/NN
Quick ML baseline	lineareg, randomforest	Fast training, easy to interpret
Tabular data, many features	catboost, lightgbm, xgboost	Strong with heterogeneous features
GPU acceleration needed	catboost (GPU), xgboost (CUDA), mlpf, nhits, rnn	Native GPU support
Parametric uncertainty — symmetric	gausscatboost, ngboostnormal, mlpfnormal, nhitsnormal, rnnnormal	Gaussian output with mean + sigma
Parametric uncertainty — heavy tails	mlpgamstudentt, mlpgafstudentt, nhitsstudentt, rnnstudentt	Student-t handles heavier-tailed residuals
Parametric uncertainty — positive targets	ngboostlognormal, mlpflognormal, mlpgamgamma, rnngamma	LogNormal / Gamma for count or power data
Distribution-free quantiles — ML	qrcatboost, qrxgboost, qrlightgbm, qrrandomforest	Pinball loss, any quantile at inference
Distribution-free quantiles — NN	mlpfqr, mlpgamqr, mlpgafqr, nhitsqr, rnnqr	Simultaneous multi-quantile output
Learned quantile levels	mlpffpqr, mlpgamfpqr, mlpgaffpqr, nhitsfpqr, rnnfpqr	FPQR adapts quantile levels during training
Post-hoc coverage guarantee	mlpfcrc, mlpgamcrc, mlpgafcrc, nhitscrc	Conformal residual coverage wraps any NN
Sequential / recurrent patterns	rnn	GRU/LSTM captures step-by-step temporal dependencies
Complex temporal patterns	nhits, mlpf	Deep learning captures nonlinearities
Interpretable forecasts	mlpgam, mlpgaf	Additive structure provides feature insights
Small datasets	lineareg, catboost, randomforest	Less prone to overfitting

Model Registry

Models are loaded dynamically via the registry in twiga/forecaster/registry.py. The naming convention is:

• Module: twiga.models.{domain}.{name}_model

• Model class: {NAME}Model (uppercase)

• Config class: {NAME}Config (uppercase)

For example, get_model("catboost", "ml") loads from twiga.models.ml.catboost_model and returns (CATBOOSTModel, CATBOOSTConfig).

See Creating Custom Models for how to add new models to the registry.