GAM timeseries modeling with auto-changepoint detection. Inspired by Facebook Prophet and implemented in PyMC3

Overview

pm-prophet

Logo

Pymc3-based universal time series prediction and decomposition library (inspired by Facebook Prophet). However, while Faceook prophet is a well-defined model, pm-prophet allows for total flexibility in the choice of priors and thus is potentially suited for a wider class of estimation problems.

⚠️ Only supports Python 3

Table of Contents

Installing pm-prophet

PM-Prophet installation is straightforward using pip: pip install pmprophet

Note that the key dependency of pm-prophet is PyMc3 a library that depends on Theano.

Key Features

  • Nowcasting & Forecasting
  • Intercept, growth
  • Regressors
  • Holidays
  • Additive & multiplicative seasonality
  • Fitting and plotting
  • Custom choice of priors (not in Facebook's prophet original model)
  • Changepoints in growth
  • Automatic changepoint location detection (not in Facebook's prophet original model)
  • Fitting with NUTS/AVDI/Metropolis

Experimental warning ⚠️

  • Note that automatic changepoint detection is experimental

Differences with Prophet:

  • Saturating growth is not implemented
  • Uncertainty estimation is different
  • All components (including seasonality) need to be explicitly added to the model
  • By design pm-prophet places a big emphasis on posteriors and uncertainty estimates, and therefore it won't use MAP for it's estimates.
  • While Faceook prophet is a well-defined fixed model, pm-prophet allows for total flexibility in the choice of priors and thus is potentially suited for a wider class of estimation problems

Peyton Manning example

Predicting the Peyton Manning timeseries:

import pandas as pd
from pmprophet.model import PMProphet, Sampler

df = pd.read_csv("examples/example_wp_log_peyton_manning.csv")
df = df.head(180)

# Fit both growth and intercept
m = PMProphet(df, growth=True, intercept=True, n_changepoints=25, changepoints_prior_scale=.01, name='model')

# Add monthly seasonality (order: 3)
m.add_seasonality(seasonality=30, fourier_order=3)

# Add weekly seasonality (order: 3)
m.add_seasonality(seasonality=7, fourier_order=3)

# Fit the model (using NUTS)
m.fit(method=Sampler.NUTS)

ddf = m.predict(60, alpha=0.2, include_history=True, plot=True)
m.plot_components(
    intercept=False,
)

Model Seasonality-7 Seasonality-30 Growth Change Points

Custom Priors

One of the main reason why PMProphet was built is to allow custom priors for the modeling.

The default priors are:

Variable Prior Parameters
regressors Laplace loc:0, scale:2.5
holidays Laplace loc:0, scale:2.5
seasonality Laplace loc:0, scale:0.05
growth Laplace loc:0, scale:10
changepoints Laplace loc:0, scale:2.5
intercept Normal loc:y.mean, scale: 2 * y.std
sigma Half Cauchy tau:10

But you can change model priors by inspecting and modifying the distributions stored in

m.priors

which is a dictionary of {prior: pymc3-distribution}.

In the example below we will model an additive time-series by imposing a "positive coefficients" constraint by using an Exponential distribution instead of a Laplacian distribution for the regressors.

import pandas as pd
import numpy as np
import pymc3 as pm
from pmprophet.model import PMProphet, Sampler

n_timesteps = 100
n_regressors = 20

regressors = np.random.normal(size=(n_timesteps, n_regressors))
coeffs = np.random.exponential(size=n_regressors) + np.random.normal(size=n_regressors)
# Note that min(coeffs) could be negative due to the white noise

regressors_names = [str(i) for i in range(n_regressors)]

df = pd.DataFrame()
df['y'] = np.dot(regressors, coeffs)
df['ds'] = pd.date_range('2017-01-01', periods=n_timesteps)
for idx, regressor in enumerate(regressors_names):
    df[regressor] = regressors[:, idx]

m = PMProphet(df, growth=False, intercept=False, n_changepoints=0, name='model')

with m.model:
    # Remember to suffix _<model-name> to the custom priors
    m.priors['regressors'] = pm.Exponential('regressors_%s' % m.name, 1, shape=n_regressors)

for regressor in regressors_names:
    m.add_regressor(regressor)

m.fit(
    draws=10 ** 4,
    method=Sampler.NUTS,
)
m.plot_components()

Regressors

Automatic changepoint detection ( ⚠️ experimental)

Pm-prophet is equipped with a non-parametric truncated Dirichlet Process allowing it to automatically detect changepoints in the trend.

To enable it simply initialize the model with auto_changepoints=True as follows:

from pmprophet.model import PMProphet, Sampler
import pandas as pd

df = pd.read_csv("examples/example_wp_log_peyton_manning.csv")
df = df.head(180)
m = PMProphet(df, auto_changepoints=True, growth=True, intercept=True, name='model')
m.fit(method=Sampler.METROPOLIS, draws=2000)
m.predict(60, alpha=0.2, include_history=True, plot=True)
m.plot_components(
    intercept=False,
)

Where n_changepoints is interpreted as the truncation point for the Dirichlet Process.

Pm-prophet will then decide which changepoint values make sense and add a custom weight to them. A call to plot_components() will reveal the changepoint map:

Regressors

A few caveats exist:

  • It's slow to fit since it's a non-parametric model
  • For best results use NUTS as method
  • It will likely require more than the default number of draws to converge
Owner
Luca Giacomel
Luca Giacomel
A Tools that help Data Scientists and ML engineers train and deploy ML models.

Domino Research This repo contains projects under active development by the Domino R&D team. We build tools that help Data Scientists and ML engineers

Domino Data Lab 73 Oct 17, 2022
CinnaMon is a Python library which offers a number of tools to detect, explain, and correct data drift in a machine learning system

CinnaMon is a Python library which offers a number of tools to detect, explain, and correct data drift in a machine learning system

Zelros 67 Dec 28, 2022
Transpile trained scikit-learn estimators to C, Java, JavaScript and others.

sklearn-porter Transpile trained scikit-learn estimators to C, Java, JavaScript and others. It's recommended for limited embedded systems and critical

Darius Morawiec 1.2k Jan 05, 2023
hgboost - Hyperoptimized Gradient Boosting

hgboost is short for Hyperoptimized Gradient Boosting and is a python package for hyperparameter optimization for xgboost, catboost and lightboost using cross-validation, and evaluating the results o

Erdogan Taskesen 34 Jan 03, 2023
Relevance Vector Machine implementation using the scikit-learn API.

scikit-rvm scikit-rvm is a Python module implementing the Relevance Vector Machine (RVM) machine learning technique using the scikit-learn API. Quicks

James Ritchie 204 Nov 18, 2022
fMRIprep Pipeline To Machine Learning

fMRIprep Pipeline To Machine Learning(Demo) 所有配置均在config.py文件下定义 前置环境(lilab) 各个节点均安装docker,并有fmripre的镜像 可以使用conda中的base环境(相应的第三份包之后更新) 1. fmriprep scr

Alien 3 Mar 08, 2022
A benchmark of data-centric tasks from across the machine learning lifecycle.

A benchmark of data-centric tasks from across the machine learning lifecycle.

61 Dec 28, 2022
Predicting diabetes over a five year period using logistic regression and the Pima First-Nation dataset

Diabetes This script uses the Pima First Nations dataset to create a model to predict whether or not an individual will develop Diabetes Mellitus Type

1 Mar 28, 2022
Predict the income for each percentile of the population (Python) - FRENCH

05.income-prediction Predict the income for each percentile of the population (Python) - FRENCH Effectuez une prédiction de revenus Prérequis Pour ce

1 Feb 13, 2022
Winning solution for the Galaxy Challenge on Kaggle

Winning solution for the Galaxy Challenge on Kaggle

Sander Dieleman 483 Jan 02, 2023
Probabilistic programming framework that facilitates objective model selection for time-varying parameter models.

Time series analysis today is an important cornerstone of quantitative science in many disciplines, including natural and life sciences as well as eco

Christoph Mark 129 Dec 24, 2022
To-Be is a machine learning challenge on CodaLab Platform about Mortality Prediction

To-Be is a machine learning challenge on CodaLab Platform about Mortality Prediction. The challenge aims to adress the problems of medical imbalanced data classification.

Marwan Mashra 1 Jan 31, 2022
moDel Agnostic Language for Exploration and eXplanation

moDel Agnostic Language for Exploration and eXplanation Overview Unverified black box model is the path to the failure. Opaqueness leads to distrust.

Model Oriented 1.2k Jan 04, 2023
A Python implementation of the Robotics Toolbox for MATLAB

Robotics Toolbox for Python A Python implementation of the Robotics Toolbox for MATLAB® GitHub repository Documentation Wiki (examples and details) Sy

Peter Corke 1.2k Jan 07, 2023
CyLP is a Python interface to COIN-OR’s Linear and mixed-integer program solvers (CLP, CBC, and CGL)

CyLP CyLP is a Python interface to COIN-OR’s Linear and mixed-integer program solvers (CLP, CBC, and CGL). CyLP’s unique feature is that you can use i

COIN-OR Foundation 161 Dec 14, 2022
Class-imbalanced / Long-tailed ensemble learning in Python. Modular, flexible, and extensible

IMBENS: Class-imbalanced Ensemble Learning in Python Language: English | Chinese/中文 Links: Documentation | Gallery | PyPI | Changelog | Source | Downl

Zhining Liu 176 Jan 04, 2023
Implementation of the Object Relation Transformer for Image Captioning

Object Relation Transformer This is a PyTorch implementation of the Object Relation Transformer published in NeurIPS 2019. You can find the paper here

Yahoo 158 Dec 24, 2022
The Emergence of Individuality

The Emergence of Individuality

16 Jul 20, 2022
A demo project to elaborate how Machine Learn Models are deployed on production using Flask API

This is a salary prediction website developed with the help of machine learning, this makes prediction of salary on basis of few parameters like interview score, experience test score.

1 Feb 10, 2022
Lingtrain Alignment Studio is an ML based app for texts alignment on different languages.

Lingtrain Alignment Studio Intro Lingtrain Alignment Studio is the ML based app for accurate texts alignment on different languages. Extracts parallel

Sergei Averkiev 186 Jan 03, 2023