What is PyFlux?¶
PyFlux is a library for time series analysis and prediction. Users can choose from a flexible range of modelling and inference options, and use the output for forecasting and retrospection. Users can build a full probabilistic model where the data \(y\) and latent variables (parameters) \(z\) are treated as random variables through a joint probability \(p\left(y,z\right)\). The advantage of a probabilistic approach is that it gives a more complete picture of uncertainty, which is important for time series tasks such as forecasting. Alternatively, for speed, users can simply use Maximum Likelihood estimation for speed within the same unified API.
The latest release version of PyFlux is available on PyPi. Python 2.7 and Python 3.5 are supported, but development occurs primarily on 3.5. To install
pyflux, simply call
pip install pyflux
PyFlux requires a number of dependencies, in particular
The development code can be accessed on GitHub. The project is open source, and contributions are welcome and encouraged. We also encourage you to check out other modelling libraries written in Python including
The PyFlux API is designed to be as clear and concise as possible, meaning it takes a minimal number of steps to conduct the model building process. The high-level outline is detailed below.
The first step is to create a model instance, where the main arguments are (i) a data input, such as a pandas dataframe, (ii) design parameters, such as autoregressive lags for an ARIMA model, and (iii) a family, which specifies the distribution of the modelled time series, such as a Normal distribution.
my_model = pf.ARIMA(data=my_dataframe, ar=2, ma=0, family=pf.Normal())
The second step is prior formation, which involves specifying a family for each latent variable in the model using the
adjust_prior method, for example we can a prior for the constant in the ARIMA model \(N\left(0,10\right)\). The latent variables can be viewed by printing the
latent_variables object attached to the model. Prior formation be ignored if the user is intending to just do Maximum Likelihood.
print(my_model.latent_variables) Index Latent Variable Prior Prior Hyperparameters V.I. Dist Transform ======== =================== =============== ========================= ========== ========== 0 Constant Normal mu0: 0, sigma0: 3 Normal None 1 AR(1) Normal mu0: 0, sigma0: 0.5 Normal None 2 AR(2) Normal mu0: 0, sigma0: 0.5 Normal None 3 Normal Scale Flat n/a (noninformative) Normal exp my_model.adjust_prior(0, pf.Normal(0, 10)) print(my_model.latent_variables) Index Latent Variable Prior Prior Hyperparameters V.I. Dist Transform ======== =================== =============== ========================= ========== ========== 0 Constant Normal mu0: 0, sigma0: 10 Normal None 1 AR(1) Normal mu0: 0, sigma0: 0.5 Normal None 2 AR(2) Normal mu0: 0, sigma0: 0.5 Normal None 3 Normal Scale Flat n/a (noninformative) Normal exp
The third step is model fitting (or inference), which involves using a
fit method, specifying an inference option. Current options include Maximum Likelihood (MLE), Metropolis-Hastings (M-H), and black box variational inference (BBVI). Once complete, the model latent variable information will be updated, and the user can proceed to the post fitting methods.
x = my_model.fit('M-H') Tuning complete! Now sampling. Acceptance rate of Metropolis-Hastings is 0.2915 x.summary() Normal ARIMA(2,0,0) ======================================== ================================================== Dependent Variable: sunspot.year Method: Metropolis Hastings Start Date: 1702 Unnormalized Log Posterior: -1219.7028 End Date: 1988 AIC: 2447.40563132 Number of observations: 287 BIC: 2462.04356018 =========================================================================================== Latent Variable Median Mean 95% Credibility Interval ========================= ================== ================== ========================= Constant 14.6129 14.5537 (11.8099 | 17.1807) AR(1) 1.3790 1.3796 (1.3105 | 1.4517) AR(2) -0.6762 -0.6774 (-0.7484 | -0.6072) Normal Scale 16.6720 16.6551 (15.5171 | 17.8696) ===========================================================================================
The fourth step is model evaluation, retrospection and prediction. Once the model has been fit, the user can look at historical fit, criticize with posterior predictive checks, predict out of sample, and perform a range of other tasks for their model.
# Some example tasks my_model.plot_fit() # plots the fit of the model my_model.plot_sample(nsims=10) # draws samples from the model my_model.plot_ppc(T=np.mean) # plots histogram of posterior predictive check for mean my_model.plot_predict(h=5) # plots predictions for next 5 time steps my_model.plot_predict_is(h=5) # plots rolling in-sample prediction for past 5 time steps predictions = my_model.predict(h=5, intervals=True) # outputs dataframe of predictions samples = my_model.sample(nsims=10) # returns 10 samples from the data ppc_pvalue = my_model.ppc(T=np.mean) # p-value for mean posterior predictive test
Want to learn how to use this library? Check out these tutorials:
- ARIMA models
- ARIMAX models
- DAR models
- Dynamic Linear regression models
- Beta-t-EGARCH models
- Beta-t-EGARCH in-mean models
- Beta-t-EGARCH in-mean regression models
- Beta-t-EGARCH long memory models
- Beta Skew-t GARCH models
- Beta Skew-t in-mean GARCH models
- GARCH models
- GAS models
- GAS local level models
- GAS local linear trend models
- GAS ranking models
- GAS regression models
- GASX models
- GP-NARX models
- Gaussian Local Level models
- Gaussian Local Linear Trend models
- Non-Gaussian Dynamic Regression Models
- Non-Gaussian Local Level Models
- Non-Gaussian Local Linear Trend Models
- VAR models