Quick Start Guide#

This guide walks through a complete forecasting workflow: loading data, configuring models, training, predicting, and evaluating.

Minimal Example with ML Models#

import pandas as pd
from sklearn.preprocessing import RobustScaler, StandardScaler

from twiga.core.config import DataPipelineConfig, ForecasterConfig
from twiga.forecaster.core import TwigaForecaster
from twiga.models.ml.xgboost_model import XGBOOSTConfig
from twiga.models.ml.lightgbm_model import LIGHTGBMConfig

# 1. Load your time series data
data = pd.read_parquet("data/my_timeseries.parquet")
# Expected columns: timestamp, target_variable, [optional features]

# 2. Configure the data pipeline
data_config = DataPipelineConfig(
    target_feature="load_mw",
    period="30min",
    lookback_window_size=48,
    forecast_horizon=48,
    calendar_features=["hour", "day_night"],
    exogenous_features=["ghi"],
    latitude=-6.8,
    longitude=39.3,
    input_scaler=StandardScaler(),
    target_scaler=RobustScaler(),
)

# 3. Configure models
xgb_config = XGBOOSTConfig(device="cpu")
lgb_config = LIGHTGBMConfig()

# 4. Configure training parameters
train_config = ForecasterConfig(
    split_freq="days",
    train_size=14,
    test_size=7,
    project_name="QuickStart",
)

# 5. Create forecaster
forecaster = TwigaForecaster(
    data_params=data_config,
    model_params=[xgb_config, lgb_config],
    train_params=train_config,
)

# 6. Split data
train_df = data[data.timestamp <= "2024-06-01"]
val_df = data[(data.timestamp > "2024-05-15") & (data.timestamp <= "2024-07-01")]
test_df = data[data.timestamp > "2024-07-01"]

# 7. Fit models
forecaster.fit(train_df=train_df, val_df=val_df)

# 8. Evaluate point forecasts
predictions_df, metrics_df = forecaster.evaluate_point_forecast(
    test_df=test_df,
    ensemble_strategy="mean",
)

print(metrics_df.groupby("Model")[["mae", "rmse", "smape"]].mean().round(2))

Using Neural Network Models#

Neural network models use NeuralModelConfig and are created via from_data_config() for convenience:

from twiga.models.nn.mlpf_model import MLPFConfig
from twiga.models.nn.nhits_model import NHITSConfig

# Create NN configs from data config (automatically sets input/output dimensions)
mlpf_config = MLPFConfig.from_data_config(data_config)
mlpf_config.max_epochs = 20
mlpf_config.batch_size = 64

nhits_config = NHITSConfig.from_data_config(data_config)
nhits_config.max_epochs = 20

# Mix ML and NN models in a single forecaster
forecaster = TwigaForecaster(
    data_params=data_config,
    model_params=[xgb_config, lgb_config, mlpf_config, nhits_config],
    train_params=train_config,
)

forecaster.fit(train_df=train_df, val_df=val_df)
predictions_df, metrics_df = forecaster.evaluate_point_forecast(
    test_df=test_df,
    ensemble_strategy="mean",
)

Note

Neural models require the nn optional dependency group. Install with pip install twiga[nn].

Hyperparameter Tuning#

Use Optuna-powered tuning before training to find optimal hyperparameters:

# Tune all models (uses Optuna with TPE sampler and Hyperband pruner)
forecaster.tune(
    train_df=train_df,
    val_df=val_df,
    num_trials=20,
    patience=5,
)

# Then fit with tuned parameters
forecaster.fit(train_df=train_df, val_df=val_df)

See Hyperparameter Tuning for details.

Backtesting#

Evaluate model performance over multiple time-based cross-validation folds:

predictions_df, metrics_df = forecaster.backtesting(
    data=data,
    train_ratio=1.0,
    verbose=True,
    ensemble_strategy="mean",
)

# Aggregate metrics across folds
print(metrics_df.groupby(["Model", "Folds"])[["mae", "rmse"]].mean().round(2))

The backtesting parameters are controlled by ForecasterConfig:

Parameter	Description	Example
`split_freq`	Unit for split sizes	`"days"`, `"months"`
`train_size`	Training window length	`14` (14 days)
`test_size`	Test window length	`7` (7 days)
`gap`	Gap between train and test	`0`
`window`	Window strategy	`"expanding"` or `"rolling"`

See Backtesting for details.

Conformal Prediction#

Add prediction intervals using conformal prediction:

from twiga.core.config import ConformalConfig

# Configure conformal prediction
conformal_config = ConformalConfig(
    method="residual",
    score_type="res",
    alpha=0.1,  # 90% prediction interval
)

# Create forecaster with conformal prediction
forecaster = TwigaForecaster(
    data_params=data_config,
    model_params=[xgb_config],
    train_params=train_config,
    conformal_params=conformal_config,
)

# Train the model
forecaster.fit(train_df=train_df, val_df=val_df)

# Calibrate on held-out data
forecaster.calibrate(calibrate_df=val_df)

# Generate prediction intervals
interval_predictions, interval_metrics = forecaster.evaluate_interval_forecast(
    test_df=test_df,
)

See Conformal Prediction for details on available methods and score types.

Accessing Predictions#

Predictions are returned as dictionaries mapping model names to NumPy arrays:

# Point predictions
prediction_dict, inference_times = forecaster.predict(test_df=test_df)

for model_name, preds in prediction_dict.items():
    print(f"{model_name}: shape={preds.shape}")
    # Shape: (num_samples, forecast_horizon, num_targets)

# Interval predictions (after calibration)
interval_dict, inference_times = forecaster.predict_interval(test_df=test_df)

for model_name, (lower, point, upper) in interval_dict.items():
    print(f"{model_name}: lower={lower.shape}, point={point.shape}, upper={upper.shape}")

Formatting Results#

Use the built-in result formatting utilities:

from twiga.core.plot.gt import twiga_report

# Aggregate metrics
res = metrics_df.groupby("Model")[["mae", "corr", "rmse", "smape"]].mean().round(2).reset_index()
res = res.rename(columns={"mae": "MAE", "corr": "Corr", "rmse": "RMSE", "smape": "SMAPE"})

# Create a formatted comparison table
twiga_report(
    res,
    metrics=["MAE", "Corr", "SMAPE", "RMSE"],
    minimize_cols=["MAE", "SMAPE", "RMSE"],
    maximize_cols=["Corr"],
)

Next: TwigaForecaster API | Configuration System