Stochastic Processes And Observed Series

Introduction

A record of annual claim counts might show 0, 1, 0, 2, 1. Those five values are the data that were observed; they are not, by themselves, a complete description of the mechanism that could generate claim counts over time. A time-series model describes the possible values at each time and how those random values are jointly generated. The recorded sequence is one outcome of that model.

This distinction matters because the same observed sequence can be compatible with different models. A model is a probability description of possible sequences; the data are the particular sequence that occurred. The distinction is foundational for time-series analysis and aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.

Definition: A stochastic process is a collection of random variables indexed by time, written \( \{X_t:t\in T\} \). Here \(T\) is the set of time points under consideration, and each \(X_t\) is a random variable. A realization (or sample path) is the sequence of values those random variables take for one outcome of the underlying random experiment. For observations at \(t=1,\ldots,n\), one realization is written \(x_1,\ldots,x_n\), where \(x_t\) is the observed value of \(X_t\).

Process, Time Index, and Observed Values

The notation \(X_t\) represents a random variable associated with time \(t\). Before observing the process, its value is uncertain under the model. After collecting data, the recorded value at that time is denoted \(x_t\). Uppercase \(X_t\) and lowercase \(x_t\) therefore serve different roles: the former denotes a random variable; the latter denotes an observed value.

A process is not just a list of separate distributions for each time. It must specify, either directly or through assumptions, the joint probability behavior across the times being modeled. For example, knowing the probability of a claim in each year does not alone tell us whether yearly claim indicators are independent or related. That distinction affects the probability assigned to a whole observed sequence.

The time index is part of the model's organization. In a yearly process, \(t\) might range over \(1,2,\ldots\); in a process observed at specified dates, \(T\) might contain only those dates. The index identifies when a random variable applies. It does not, by itself, imply that observations are independent, identically distributed, or equally spaced.

A realization gives one value at each observed time. If the process is defined for many possible future times but data are available for only the first five, the data show only a finite segment of one realization. Other sequences could have occurred under the same model.

Worked Example: A Hypothetical Annual Claim Indicator

A hypothetical insurer models whether a specified type of claim occurs in each of four years. Let \(X_t=1\) if at least one such claim occurs in year \(t\), and \(X_t=0\) otherwise. Assume the four indicators are independent and each has a Bernoulli distribution with \(P(X_t=1)=0.25\). The observed values are \(0,1,0,1\). Identify the process and realization, then calculate the probability of this exact sequence under the model.

Method: identify the random-variable family, then use independence to obtain the path probability. The process is \(\{X_t:t=1,2,3,4\}\), where each \(X_t\) is a Bernoulli random variable with success probability \(0.25\). The observed realization is \((x_1,x_2,x_3,x_4)=(0,1,0,1)\).

For a Bernoulli variable with success probability \(0.25\), the probability of a zero is \(1-0.25=0.75\). By independence, the probability of the specified four-value path is the product of its four single-year probabilities:

$$ P(X_1=0,X_2=1,X_3=0,X_4=1) =(0.75)(0.25)(0.75)(0.25) =0.03515625. $$

Thus, the model assigns probability \(0.03515625\) to this exact ordered sequence. The answer is not the probability of any sequence with two claims: the order is specified, and this calculation concerns the one observed path. The common mistake avoided here is treating the observed values as the process itself rather than as one realization of the process.

Why a Path Probability Needs a Model

For a finite set of times, the probability of a particular realization is a joint probability. Under independence, the joint probability factors into a product of the probabilities at each time. Without independence, that product generally cannot be used; the joint distribution must account for dependence between time points.

The factorization follows from the definition of independence. For independent random variables \(X_1,\ldots,X_n\), the probability of an intersection of events, one specified for each variable, equals the product of their probabilities. Taking the events to be \(\{X_t=x_t\}\) gives the path formula for discrete-valued variables:

P(X_1=x_1,\ldots,X_n=x_n) =\prod_{t=1}^{n}P(X_t=x_t), \qquad \text{when }X_1,\ldots,X_n\text{ are independent}.

The independence condition is essential. It is not a consequence of observing values at different times. If a process has dependence across times, its path probability must be obtained from its joint distribution instead.

Worked Example: A Path of Hypothetical Claim Counts

A hypothetical portfolio model specifies independent annual claim counts \(X_1,X_2,X_3\), each with a Poisson distribution having parameter \(\lambda=2\) claims per year. In one three-year observation, the counts are \(1,0,3\). Calculate the probability of this exact realization, retaining the year order.

Method: apply the Poisson probability formula at each time, then multiply because the annual counts are specified to be independent. For a Poisson random variable with parameter \(\lambda\), \(P(X_t=k)=e^{-\lambda}\lambda^k/k!\). The observed realization is \((1,0,3)\), so:

\begin{aligned} P(X_1=1,X_2=0,X_3=3) &=P(X_1=1)P(X_2=0)P(X_3=3)\\ &=\left(e^{-2}\frac{2^1}{1!}\right) \left(e^{-2}\frac{2^0}{0!}\right) \left(e^{-2}\frac{2^3}{3!}\right)\\ &=(2e^{-2})(e^{-2})\left(\frac{8}{6}e^{-2}\right)\\ &=\frac{8}{3}e^{-6}\\ &\approx 0.0066100. \end{aligned}

Rounded to four decimal places, the path probability is \(0.0066\). The check is that the three exponential factors combine to \(e^{-6}\), while the non-exponential factors multiply to \(2\times1\times(8/6)=8/3\). The common mistake avoided is multiplying marginal probabilities without first verifying independence; independence is explicitly part of this model.

Realizations Under Discrete and Continuous Models

For discrete-valued processes, an exact finite path can have positive probability, as in the claim-count examples. For continuous-valued processes, a specified exact value at a time generally has probability zero. Consequently, an exact vector of continuous observations also generally has probability zero. This does not mean the values cannot be observed. It means probabilities are assigned to ranges of values, rather than to an exact point.

A realization should still be understood as the values produced on one outcome, whether those values are discrete or continuous. When calculating probability for a continuous process, use an event such as “the value lies between 110 and 120,” not the probability of observing exactly 112.

Worked Example: A Continuous Process With a Shared Random Level

A hypothetical model records an index at three consecutive observation times and assumes \(X_1=X_2=X_3=Z\), where the shared level \(Z\) has distribution \(N(100,25)\). Thus \(Z\) is normal with mean 100 and variance 25, or standard deviation 5. One observed realization is \((112,112,112)\). Find the probability that all three observations lie between 110 and 120, and explain why the exact observed triple has probability zero.

Method: use the shared-variable definition to reduce the three-time event to one normal probability. Since every \(X_t\) equals \(Z\), the event that all three values lie in the interval is exactly the event \(110<Z<120\):

P(110<X_1<120,\ 110<X_2<120,\ 110<X_3<120) =P(110<Z<120).

Standardize \(Z\) using its mean 100 and standard deviation 5. The lower endpoint becomes \((110-100)/5=2\), and the upper endpoint becomes \((120-100)/5=4\). If \(\Phi\) is the standard normal cumulative distribution function, then:

P(110<Z<120)=\Phi(4)-\Phi(2) \approx 0.99997-0.97725 \approx 0.0227.

The three observations move together because they are all the same random variable \(Z\); this is not a model of three independent normal values. The exact triple \((112,112,112)\) corresponds to the event \(Z=112\), which has probability zero because a continuous normal distribution assigns probability zero to any single exact value. The interval event has positive probability. This avoids the common mistake of multiplying three separate interval probabilities as if the observations were independent.

What One Observed Series Can Establish

A recorded series is evidence about a process, but it does not uniquely determine the process. A sequence such as \(0,1,0,1\) may be possible under several models with different assumptions. To calculate probabilities or make forecasts, an analyst needs a model that specifies how the random variables are generated, including relevant assumptions about their joint behavior.

Keep these distinctions explicit when reading or describing a time-series problem:

Process: the full collection of random variables indexed by the times of interest.
Realization: the values generated by that process for one outcome.
Observed series: the recorded portion of a realization, potentially limited to a finite number of times.
Model assumptions: the probability rules used to describe possible values and their joint behavior.

A frequent exam trap is to infer a model property from the appearance of a single path. For instance, alternating observed values do not by themselves prove that future values must alternate, nor do a few similar values prove that observations are independent or identical in distribution. Those are claims about the process and require assumptions or evidence beyond merely naming the observed realization.

Key takeaway: \( \{X_t:t\in T\} \) names the random process; \(x_1,\ldots,x_n\) records one observed realization. A probability for an entire path requires a joint model, and multiplication of separate time-specific probabilities is justified only when the required independence assumption holds.

Check Your Understanding

For each problem, distinguish the modeled process from the observed path and use only the assumptions stated.

A hypothetical annual indicator process has independent Bernoulli variables with success probability \(0.4\). The observed four-year path is \((0,1,0,1)\). What probability does the model assign to this exact ordered path? (Answer: \(0.0576\))
A hypothetical process consists of two independent Poisson claim counts, each with parameter \(\lambda=1.5\). For the observed path \((2,0)\), calculate its probability to four decimal places. (Answer: \(0.0560\))
For a hypothetical three-time process, \(X_1=X_2=X_3=Z\), with \(Z\sim N(100,25)\). What is the probability that all three observations are between 95 and 105? (Answer: \(0.6827\))
A hypothetical process has independent annual Bernoulli claim indicators with probability \(0.10\) of a claim in each year. What is the probability of observing no claims in five specified years? (Answer: \(0.59049\))
A model says \(X_1\) and \(X_2\) each have probability \(0.3\) of being 1, and the observed path is \((1,1)\). Is the path probability necessarily \(0.09\)? State the additional condition that would make that calculation valid. (Answer: No; it is \(0.09\) only if \(X_1\) and \(X_2\) are independent.)

What You'll Learn