What are the usual approach to modelling binary time series? Is there a paper or a text book where this is treated? I think of a binary process with strong auto-correlation. Something like the sign of an AR(1) process starting at zero. Say $X_0 = 0$ and $$ X_{t+1} = \beta_1 X_t + \epsilon_t, $$ with white noise $\epsilon_t$. Then the binary time series $(Y_t)_{t \ge 0}$ defined by $$ Y_t = \text{sign}(X_t) $$ will show autocorrelation, which I would like to illustrate with the following code

X = rep(0,100)
beta = 0.9
sigma = 0.1
for(i in 1:(length(X)-1)){
  X[i+1] =beta*X[i] + rnorm(1,sd=sigma)

What is the text book/usual modelling approach if I get the binary data $Y_t$ and all I know is that there is significant autocorrelation?

I thought that in case of external regressors or seasonal dummies given I can do a logistic regression. But what is the pure time-series approach?

Plot of the ACF of the sign

EDIT: to be precise let's assume that sign(X) is autocorrelated for up to 4 lags. Would this be a Markov model of order 4 and can we do fitting and forecasting with it?

EDIT 2: In the meanwhile I stumbled upon time series glms. These are glms where the explanatory variables are lagged observations and external regressors. However it seems that this is done for Poisson and negative binomial distributed counts. I could approximate the Bernoullis using a Poisson distribution. I just wonder whether there is no clear text book approach to this.

EDIT 3: bounty expires ... any ideas?

  • $\begingroup$ For your specific example you could try using a usual ar process as a latent process, only observing the indicator, and then set up the likelihood function. $\endgroup$ – kjetil b halvorsen Mar 14 '16 at 21:14
  • $\begingroup$ This would be one way to go ... but what if O don't know where the binary process comes form? Then the above would bear a lot of model risk. Please see my edit for more info. $\endgroup$ – Ric Mar 15 '16 at 7:03
  • 1
    $\begingroup$ You might want to try searching dimer models. These are similar. Here is a paper that might be useful arxiv.org/pdf/1406.2656.pdf. $\endgroup$ – Greg Petersen Mar 17 '16 at 21:05
  • 1
    $\begingroup$ See robjhyndman.com/papers/ar1.pdf $\endgroup$ – Yves Mar 21 '16 at 10:34
  • 1
    $\begingroup$ A reference for binary variate in the former article is available as researchgate.net/publication/… ' section 4.6. Sorry no package reference, and I might lack of time for an answer. $\endgroup$ – Yves Mar 21 '16 at 13:09

The class of score-driven models might be of interest to you:

Creal, D. D., S. J. Koopman, and A. Lucas (2013). Generalized autoregressive score models with applications. Journal of Applied Econometrics 28(5), 777--795

Score-driven models were applied to a binary time series of outcomes of The Boat Race between Oxford and Cambridge,


In that paper, the time-varying probability was obtained with the score-driven methodology by using the (free) Time Series Lab software package.

The score-driven model for binary observations in short:

Our observations can take on either two values: 0 and 1. We therefore assume that these observations come from the Binary distribution with probability density function (pdf) \begin{equation}\label{eq:pdf} p(y_t | \pi_t) = \pi_t^{y_t} (1-\pi_t)^{1-y_t} \end{equation} where $\pi_t$ is a time-varying probability and $y_t \in \{0,1\}$ for $t = 1,\ldots,T$ where $T$ is the length of the time series. We can specify the time-varying probability $\pi_t$ as a function of a dynamic process $\alpha _t$, that is \begin{equation}\label{eq:pi} \begin{aligned} \pi_{t+1} &= f(\alpha_t),\\ \alpha_{t+1} &= \omega + \phi \alpha_t + \kappa s_t , \end{aligned} \end{equation} where the link function $f(\cdot)$ is the logit link function so that $\pi_t$ takes values between 0 and 1. You can easily include more lags of $\alpha_t$ or $s_t$ into the equation above. The unknown coefficients include the constant $\omega$, the autoregressive parameter $\phi$, and the score updating parameter $\kappa$ which are estimated by maximum likelihood. The driving force behind the updating equation of $\alpha_t$ is the scaled score innovation $s_t$ as given by \begin{equation}\label{eq:score} \qquad s_t = S_t \cdot \nabla_t, \qquad \nabla_t = \frac{\partial \, \log \, p(y_t | \pi_t)}{\partial \alpha_t}, \end{equation} for $t = 1,\ldots,T$ and where $\nabla_{t}$ is the score of the density $p(y_t | \pi_t)$ and $S_t$ a scaling factor which is often the inverse of the Fisher information.

[Disclaimer] I am one of the developers of Time Series Lab.

| cite | improve this answer | |
  • $\begingroup$ Thank you for this very interesting approach and the link to you page. $\endgroup$ – Ric Jun 17 at 9:26

If I understand your question correctly, the "usual approach" would be a dynamic probit approach, cf. "Predicting U.S. Recessions with Dynamic Binary Response Models", Heikki Kauppi and Pentti Saikkonen, The Review of Economics and Statistics Vol. 90, No. 4 (Nov., 2008), pp. 777-791, The MIT Press, Stable URL: http://www.jstor.org/stable/40043114

Whether that model class directly reflects your underlying example process might depend on what epsilon_t is like exactly, but I think the model fits your statement "all I know is that there is significant autocorrelation".

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.