Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Suppose you have recorded a set of paths in the $y,t$ plane, with $y = f(t)$, $f$ is a stochastic function (i.e. there is a noise term), and $t$ might be time or some other monotonic increasing independent variable.

Suppose further that you know the underlying process is random, meaning the same set of initial conditions (even, "hidden variables") do not produce the same path. Do not assume that changes in $y$ are independent, or even stationary (though, stationary but dependent changes are an important special case).

What techniques are available to simulate an incomplete curve into the future?
What I mean by "incomplete" curve is: your set of historical curves are each over some range of interest, $T=[0,t_{max}]$. You are given a curve that represents data $y_1(t)$ recorded up until some intermediate value of $t' \in T$.

share|improve this question
Are you willing to assume that the historical data form a set of identically distributed and perhaps even independent curves? – NRH Aug 12 '11 at 20:07
Yes, good point; an important case is with each curve independent from the other, with the same underlying distribution throughout $T$. However, it is important that changes in $y$ on a given curve are generally dependent. – Pete Aug 12 '11 at 23:38
Coudl you use Markov Chains? – Manoel Galdino Aug 16 '11 at 2:08
up vote 3 down vote accepted

Would dynamic linear models be applicable? (State Space formulation, Kalman filter, etc.) The dlm package has some nice tools to create and simulate from models.

share|improve this answer

Conceptually, the question fits into the framework of functional data analysis, see, for instance Applied Functional Data Analysis by Ramsay and Silverman. The usual assumption here is that we have a data set of independent, and perhaps even identically distributed, smooth curves. Fitting an fda model to your data over $[0, t_{\max}]$ you are able to predict a future $y_1(t)$ based on the fitted model, or, in principle, the continuation of $y_1(t)$ for $t > t'$ based on the conditional distribution from the fitted model of $y_1(t)$ for $t > t'$ given $y_1(t)$ for $0 \leq t \leq t'$. However, this may be easier said than done.

A simple example is when your curves can be modeled via an ordinary differential equation (with random initial values). Then you can predict $y_1(t)$ for $t > t'$ by solving the ode with the observed initial condition $y_1(t')$.

I would recommend that you take a look in the book above, and perhaps also their theoretical version Functional data analysis or their web page for inspiration. I think there are many ways to proceed depending on the relative impact of inhomogeneity in the $t$-variable and information content in the partial observation of $y_1$ on the continuation of the curve.

share|improve this answer
(+1) Good suggestions. Even the "theoretical version" is an easy and enjoyable read. :) – cardinal Oct 1 '11 at 23:35

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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