Skip to content

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

Notifications You must be signed in to change notification settings

luke14free/pm-prophet

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

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

About

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

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages