Define the clock and matrix orientation
A transition matrix describes movements between states over a specified interval. Declare whether that interval is a session, a release, or another observation step. In this guide, rows are origin states and columns are destination states; other conventions exist. A descriptive count matrix can be constructed without claiming that the labels follow a Markov process. Using it as a forecasting model adds assumptions, including whether the current state contains the relevant history and whether transition behavior remains stable. Gallager's finite-state Markov-chain notes explain this model structure.
Daily and monthly self-transition fractions are not directly comparable measures of persistence. A longer interval can contain departures and returns that disappear between endpoints. Matrix powers describe longer-step probabilities only under the specified transition model; changing the labels' sampling frequency is not merely a cosmetic change to the table.
Count eligible adjacent pairs, not just labels
A state occurrence contributes an outgoing transition only when the next eligible observation exists. The final observation therefore has no observed outgoing move. Missing intermediate observations should not silently become direct one-step transitions across the gap. Record skipped pairs and distinguish an unknown label from an observed state. Likewise, repeating a monthly input on every daily row can produce many self-transitions without adding new releases. Choose the clock for the question being studied, and preserve that choice when comparing matrices across versions or samples.
Normalize counts without hiding denominators
For each origin state, divide each destination count by the number of eligible outgoing pairs from that origin. Display those raw counts beside the resulting fractions. A row with no eligible pairs is undefined, not a row of zero probabilities. A zero count in an otherwise populated row means that a transition did not appear in that sample; it does not establish impossibility. Pseudocounts can prevent extreme estimates, but they introduce a declared prior or regularization choice, not additional observed transitions. Keep adjusted and raw rows separate.
Work through an invented short sequence
The entirely invented sequence A, A, B, B, A, A, A, B, B, A contains ten labels and nine adjacent pairs. There are three A-to-A moves, two A-to-B moves, two B-to-A moves, and two B-to-B moves. Consequently A has five outgoing pairs and B has four. Their descriptive rows are 60%/40% and 50%/50%. These are fractions of fictional counts, not estimated real-world probabilities. The final A is excluded from the outgoing denominator because the example supplies no next observation.
Treat sparse-row uncertainty as structural
A visually precise percentage can rest on very few opportunities to leave a state. NIST's proportion-interval reference explains why small denominators require care, but a binomial interval is not automatically an exact interval for an entire Markov trajectory with a random number of visits. Estimated labels, changing transition behavior, and clustered episodes add uncertainty. Examine subperiods and distinct episodes, and state the resampling or modeling assumptions used for intervals. More decimal places cannot resolve weak coverage, and pooled data can conceal a meaningful change in the process.
Duration formulas require an extra model
Under a homogeneous first-order Markov model, a state's self-transition probability p implies geometric spell lengths with mean 1/(1-p), counting the initial occupied period. Applying that assumption to the invented A row gives 1/(1-0.6), or 2.5 observation periods. This is an assumption-dependent illustration, not an observed average spell or a forecast. A near-one estimate makes duration especially sensitive to error. A sample with no exits does not prove an absorbing state; it may simply be too short to contain a departure.
Keep label uncertainty outside the count arithmetic
Counts from deterministic labels describe that classification rule. Counts from a fitted model's most likely labels discard its state uncertainty and are not generally its fitted transition parameters. Filtered and smoothed labels also use different information sets. Archive the label method, parameters, vintage policy, clock, and boundary handling with every matrix. Hysteresis can make a label depend on prior state, so a compact matrix may omit important memory. This guide is a descriptive research framework, not investment advice or evidence that a real market will follow the invented transitions.
Inspect the comparison
| Invented origin | Count to A | Count to B | Outgoing pairs | Fraction to A | Fraction to B |
|---|---|---|---|---|---|
| A | 3 | 2 | 5 | 60% | 40% |
| B | 2 | 2 | 4 | 50% | 50% |
Primary references
References support the stated methods and data conventions. Original fictional examples are not observations from the cited sources.