PASTS — Python AnalySis for Time Series

A unified interface for time series analysis and forecasting. All models — parametric (e.g. linear trend), ML/DL (via Darts), and ensembles — share the same TimeSeriesModel interface.

Features

• Statistical testing — stationarity, seasonality, normality, white noise, heteroscedasticity, Granger causality
• Decomposition — imperative workflow with automatic recomposition: trend removal, scaling, arbitrary transforms
• Visualization — signal.plot accessor with matplotlib and Plotly backends
• Pluggable models — LinearTrend, MovingAverageTrend, HPFilterTrend, STLTrend, EMDTrend, HighPassFilterTrend, Differencing, DartsModel, AggregatedModel

Installation

pip install git+https://github.com/eurobios-mews-labs/pasts

Architecture

Signal (extends DataCube)
├── TimeSeriesModel (ABC)            — unified fit / transform / reverse_transform
│   ├── LinearTrend                  — linear parametric model
│   ├── NonParametricTrend           — MovingAverage, HPFilter, STL, EMD, HighPassFilter
│   ├── DartsModel                   — wraps any Darts ML/DL estimator
│   └── AggregatedModel              — RMSE-weighted ensemble
├── Decomposition                    — records and reverses signal transformations
├── ValidationAccessor (.validation) — train/test splits, optional cross-validation
├── StatAccessor (.stat)             — statistical tests per column
├── Metrics                          — R², MSE, RMSE, MAPE, SMAPE, MAE
└── PlotAccessor (.plot)             — matplotlib / Plotly plots

Typical workflow:

1. signal.stat.test_*() — run statistical tests
2. signal.validation_split() — define train/test boundary
3. signal.decompose() + signal.decompositions["name"].apply_model() — decompose and fit models
4. signal.forecast() — predict through decomposition back to original space
5. signal.compute_scores() — evaluate metrics on held-out set
6. signal.apply_model(AggregatedModel(...)) — combine models weighted by RMSE

Full example: examples/readme/model.py

Quick start

Load data

import pandas as pd
from darts.datasets import ETTh1Dataset
from pasts.signal import Signal

# Electricity Transformer Temperature — 7 variables (6 power loads + oil temperature)
# Hourly data from 2016-07 to 2018-06 (~17 400 observations)
series = ETTh1Dataset().load()
df_all = pd.DataFrame(series.values())
df_all.columns = ['HUFL', 'HULL', 'MUFL', 'MULL', 'LUFL', 'LULL', 'OT']
df_all.index = series.time_index

# Resample to daily frequency for manageable computation
df_daily = df_all.resample('D').mean()

# Focus on Oil Temperature (OT), the main target variable
dt = df_daily[['OT']]
signal = Signal(dt, path='examples/readme/ETTh1_univariate')

Statistical tests

All tests return a pd.DataFrame and are applied per-column on multivariate data. Results are also stored in signal.tests_stat.

signal.stat.test_stationarity()
signal.stat.test_stationarity(method='kpss')
signal.stat.test_seasonality()

Visualization

signal.plot()        # raw signal
signal.plot.acf()    # autocorrelation function

Train/test split

timestamp = '2018-04-01'
signal.validation_split(timestamp=timestamp)

signal.train_data   # pd.DataFrame — computed on demand
signal.test_data    # pd.DataFrame — computed on demand

signal.validation.split(timestamp) is the accessor form; both are equivalent.

Decomposition

Named decompositions allow trend removal while keeping everything linked to the parent signal. The split defined on the parent is automatically shared with each decomposition.

from pasts.components import MovingAverageTrend

signal.decompose("MA_Trend")
signal.decompositions["MA_Trend"] -= MovingAverageTrend(30)

Each operation is recorded symbolically; compose() walks the stack in reverse to reconstruct the original signal from a predicted residual.

Available trend components:

Component	Description
LinearTrend	Linear parametric model
MovingAverageTrend	Moving average filter
HPFilterTrend	Hodrick-Prescott filter
STLTrend	STL decomposition
EMDTrend	Empirical Mode Decomposition
HighPassFilterTrend	High-pass frequency filter
Differencing	Transform-based detrending via DataCube.apply()

Fitting models

Models are applied directly on a named decomposition. save_model=True persists the fitted estimator and its predictions to disk (joblib, under signal.path).

from darts.models import XGBModel, RandomForestModel

signal.decompositions["MA_Trend"].apply_model(XGBModel(lags=250), save_model=True)
signal.decompositions["MA_Trend"].apply_model(RandomForestModel(lags=250), save_model=True)

Forecast

forecast() composes predictions back to the original signal space through the decomposition. signal.models["MA_Trend__XGBModel"] is populated with the composed predictions afterwards.

signal.forecast("MA_Trend__XGBModel", 100, save_model=True)
signal.forecast("MA_Trend__RandomForestModel", 100, save_model=True)

Scores

signal.compute_scores()

Available metrics: r2, mse, rmse, mape, smape, mae.

signal.plot.predictions()    # predictions vs actual

Aggregated model

AggregatedModel combines models weighted by their RMSE on test data.

from pasts.components.aggregated_model import AggregatedModel

signal.apply_model(AggregatedModel(
    {'MA_Trend__XGBModel': signal.models['MA_Trend__XGBModel']['model'],
     'MA_Trend__RandomForestModel': signal.models['MA_Trend__RandomForestModel']['model']},
), save_model=True)
signal.compute_scores(axis=1)
signal.compute_conf_intervals()
signal.forecast("AggregatedModel", 100, save_model=True)

signal.plot.forecast()                        # all models
signal.plot.forecast(aggregated_only=True)    # aggregated model only

Multivariate

The same workflow applies to multivariate data. test_causality() is available for multivariate signals.

df_m = df_daily[['HUFL', 'MUFL', 'LUFL', 'OT']]
signal_m = Signal(df_m, path='examples/readme/ETTh1_multivariate')

signal_m.stat.test_causality()

timestamp = '2018-04-01'
signal_m.validation_split(timestamp=timestamp)

signal_m.decompose("MA_Trend")
signal_m.decompositions["MA_Trend"] -= MovingAverageTrend(30)

lags = [-325]
signal_m.decompositions["MA_Trend"].apply_model(XGBModel(lags=lags), save_model=True)
signal_m.decompositions["MA_Trend"].apply_model(RandomForestModel(lags=lags), save_model=True)

signal_m.forecast("MA_Trend__XGBModel", 50, save_model=True)
signal_m.forecast("MA_Trend__RandomForestModel", 50, save_model=True)
signal_m.compute_scores(axis=1)

signal_m.apply_model(AggregatedModel(
    {'MA_Trend__XGBModel': signal_m.models['MA_Trend__XGBModel']['model'],
     'MA_Trend__RandomForestModel': signal_m.models['MA_Trend__RandomForestModel']['model']},
), save_model=True)
signal_m.compute_scores()

signal_m.forecast("AggregatedModel", 50, save_model=True)

Documentation

pip install sphinx sphinx_rtd_theme
cd doc && make html

Then open doc/_build/html/index.html.

---