Pipeline

The ForecastPipeline chains preprocessing transformers with a forecasting model. Transformations are automatically inverted when generating predictions, so your forecasts come back in the original data scale.

Basic Pipeline

from vectrix.pipeline import ForecastPipeline, Scaler, OutlierClipper

pipe = ForecastPipeline([
    ('clip', OutlierClipper()),
    ('scale', Scaler()),
    ('forecast', MyForecaster()),
])

pipe.fit(y)
predictions, lower, upper = pipe.predict(12)

The pipeline

fit — sequentially fitTransform each transformer, then fit the forecaster on transformed data
predict — get predictions from the forecaster, then inverseTransform through all transformers in reverse order

Built-in Transformers

Transformer	What it does	Inverse?
`Scaler(method='zscore')`	Z-score standardization or MinMax normalization	Yes
`LogTransformer()`	`log(1 + y)` with automatic shift for negative values	Yes
`BoxCoxTransformer()`	Optimal Box-Cox lambda estimation via MLE	Yes
`Differencer(d=1)`	d-th order differencing for stationarity	Yes
`Deseasonalizer(period=7)`	Remove seasonal component by period averaging	Yes
`Detrend()`	Remove linear trend	Yes
`OutlierClipper(factor=3.0)`	IQR-based outlier clipping	No
`MissingValueImputer(method='linear')`	Fill NaN via linear interpolation, mean, or forward fill	No

Multi-step Preprocessing

Chain multiple transformers for complex data preparation

from vectrix.pipeline import (
    ForecastPipeline, MissingValueImputer, OutlierClipper,
    LogTransformer, Deseasonalizer, Scaler
)

pipe = ForecastPipeline([
    ('impute', MissingValueImputer(method='linear')),
    ('clip', OutlierClipper(factor=3.0)),
    ('log', LogTransformer()),
    ('deseason', Deseasonalizer(period=7)),
    ('scale', Scaler(method='zscore')),
    ('forecast', MyForecaster()),
])

Transform Without Forecasting

Use the pipeline as a pure preprocessing tool

pipe.fit(train_data)

transformed = pipe.transform(test_data)
original = pipe.inverseTransform(transformed)

Inspecting the Pipeline

pipe.listSteps()
# ['impute', 'clip', 'log', 'deseason', 'scale', 'forecast']

scaler = pipe.getStep('scale')
print(scaler._mean, scaler._std)

pipe.getParams()
# {'clip__factor': 3.0, 'scale__method': 'zscore', ...}

Scaler Options

Z-score (default)

Scaler(method='zscore')

Centers to mean=0, std=1. Best for general-purpose standardization.

MinMax

Scaler(method='minmax')

Scales to [0, 1] range. Best when bounded output is needed.

Box-Cox Transform

Automatically finds the optimal lambda to normalize your data distribution

from vectrix.pipeline import BoxCoxTransformer

bc = BoxCoxTransformer()       # auto-estimate lambda
bc = BoxCoxTransformer(lmbda=0.5)  # fixed lambda (square root)

Missing Value Strategies

MissingValueImputer(method='linear')  # Linear interpolation (default)
MissingValueImputer(method='mean')    # Replace with series mean
MissingValueImputer(method='ffill')   # Forward fill