Infer a changing hidden process from an observable sequence
A hidden Markov model separates an unobserved state sequence from the observations it emits, then uses probability to filter, smooth, decode or learn under explicit dependence assumptions.
Many judgement problems involve a condition that cannot be observed directly: a machine regime behind sensor readings, a phoneme behind acoustic frames, or a biological state behind measurements. Individual observations overlap, so a sequence carries more information than any single reading.
An HMM is useful when the current hidden state depends on the previous state and the current observation depends on the current state. Those assumptions create efficient dynamic-programming algorithms, but they also create characteristic failures when duration, memory, covariates or observation processes are misspecified.
Chapter contents
- 1. Define hidden states and observed emissions
- 2. Specify initial and transition probabilities
- 3. Specify emission distributions
- 4. Write the joint probability of a path and sequence
- 5. Use the forward algorithm to sum over paths
- 6. Worked example: weather and activity sequence
- 7. Use backward messages for smoothing
- 8. Decode a path with Viterbi
- 9. Learn parameters with Baum-Welch carefully
- 10. Prevent underflow and implementation errors
- 11. Diagnose state count, duration and residual dependence
- 12. Test transfer, drift and decision consequences
Choose a reading route
- Route 1: Set up the model — states, observations and probabilities
- Route 2: Compute likelihood — joint paths and forward recursion
- Route 3: Work the example — filtering, smoothing and decoding
- Route 4: Learn parameters — Baum-Welch and numerical stability
- Route 5: Test transfer — diagnostics, duration and domain shift
How to use this guide
Start with the chapters that match your immediate decision, then return to the worked example and audit chapters. Each chapter separates a model from the evidence needed to use it responsibly. A claim is not strengthened merely by being technical; it is strengthened when assumptions, observations, alternatives and limits can be inspected.
| Layer | Question | Minimum record |
|---|---|---|
| Concept | What exactly is being represented? | Definitions and boundary |
| Method | How will the reasoning be carried out? | Steps, inputs and assumptions |
| Transfer | Would it work in a changed setting? | A deliberately altered case |
| Audit | What could prove it wrong? | Evidence, limits and revision trigger |
Chapter 1. Define hidden states and observed emissions
The idea
The state is a latent condition the model uses to explain persistence and change. The observation is measured. Confusing the two produces labels that cannot be validated.
A disciplined method
Name the time index, state space, observation space and measurement process. Ask what evidence could independently check a state interpretation.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether the proposed states represent real regimes or convenient clusters. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A machine has hidden Normal and Degraded states; vibration category is observed. A loud reading is evidence, not the state itself.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Change the sensor from vibration to temperature and decide whether the same state meanings remain defensible.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Latent states may be useful abstractions without corresponding to unique physical causes.
Evidence to retain: state definitions, measurement protocol, external checks and labelling rationale
Common misconception: The most likely state is directly observed once the model is fitted.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 2. Specify initial and transition probabilities
The idea
The initial distribution describes uncertainty at the first time. The transition matrix describes how states persist or change from one step to the next. Each row forms a probability distribution.
A disciplined method
Choose a time step meaningful for the process, estimate or elicit transitions, verify row sums and inspect implied dwell times and stationary behaviour.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is what one model step represents. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
At hourly resolution Normal persists with probability 0.9 and Degraded with 0.8. At daily resolution those numbers cannot be reused unchanged.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Change sampling frequency from hourly to irregular observations and explain why a discrete homogeneous transition matrix may fail.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: First-order, time-homogeneous transitions ignore longer memory and changing covariates.
Evidence to retain: time stamps, transition counts or priors, row checks and sensitivity ranges
Common misconception: Transition probabilities are properties of the system independent of sampling interval.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 3. Specify emission distributions
The idea
The emission model links each hidden state to a distribution over observations. Overlap represents measurement noise and genuine ambiguity.
A disciplined method
Choose a distribution suited to discrete, continuous or multivariate data; fit parameters on the measurement scale; inspect overlap, tails, missingness and dependence.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which features are conditionally independent given the state. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
Normal emits Low vibration with 0.8 and High with 0.2; Degraded reverses those probabilities. A High reading is informative but not conclusive.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Add a second sensor correlated with the first and test whether multiplying independent emissions overstates certainty.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: A poor emission family can force states to absorb outliers or measurement artefacts.
Evidence to retain: raw measurements, calibration, residual plots, missing-data rules and parameter uncertainty
Common misconception: Each state must emit a unique observation.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 4. Write the joint probability of a path and sequence
The idea
An HMM factorises the probability of states and observations into an initial factor, transition factors and emission factors. This factorisation makes assumptions inspectable.
A disciplined method
For a proposed path, multiply the initial state probability, each emission probability and each transition probability in temporal order. Use log values for long sequences.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which factors belong to time t and which connect t-1 to t. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
For states N,D and observations L,H, the path N,D has probability pi(N) bN(L) aND bD(H).
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Transfer to three steps and verify that every observation contributes once and every adjacent state pair contributes once.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: The joint probability of one path is not the total likelihood, which sums over all possible paths.
Evidence to retain: factor list, indexing convention, hand calculation and software comparison
Common misconception: The probability of the best path equals the probability of the observation sequence.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 5. Use the forward algorithm to sum over paths
The idea
Forward recursion computes the probability of observations up to time t ending in each state. It reuses partial sums instead of enumerating exponentially many paths.
A disciplined method
Initialise alpha1(i)=pi(i)bi(o1); then update alphat(j)=bj(ot) sum_i alpha(t-1,i)aij. Sum final alphas for sequence likelihood.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether the task needs total likelihood rather than a single path. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
With two states and three observations, compute two alpha values per step. The total likelihood is their final sum.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Add a third state and confirm computational growth is quadratic in state count per step, not exponential in path count.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Very small probabilities underflow without scaling or log-space computation.
Evidence to retain: hand table, transition and emission matrices, scale factors and final likelihood
Common misconception: Forward recursion chooses the most likely state at each time.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
| Step | Normal alpha | Degraded alpha | Meaning |
|---|---|---|---|
| 1 | piN bN(o1) | piD bD(o1) | First observation |
| 2 | bN(o2)Σi α1(i)aiN | bD(o2)Σi α1(i)aiD | All two-step paths |
| T | αT(N) | αT(D) | Sum for likelihood |
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 6. Worked example: weather and activity sequence
The idea
A fictional HMM has Sunny and Rainy hidden weather states and observed indoor or outdoor activity. The exercise keeps numbers small enough for hand verification.
A disciplined method
Set pi=(0.6,0.4), transitions [[0.7,0.3],[0.4,0.6]], and emissions P(outdoor|Sunny)=0.8, P(outdoor|Rainy)=0.3. Compute forward values for outdoor, indoor, outdoor.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is how persistence and ambiguous emissions combine. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
The first outdoor observation favours Sunny. An indoor observation raises Rainy, but transitions preserve some Sunny probability. The final outdoor observation updates both paths rather than resetting history.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Change only the middle observation and compare the filtered and smoothed state probabilities at time two.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: The numbers are illustrative, not an empirical weather model.
Evidence to retain: all parameters, arithmetic, normalisation constants and an independent implementation
Common misconception: The current observation alone determines the current state.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 7. Use backward messages for smoothing
The idea
Backward recursion computes the likelihood of future observations given a state at time t. Combining forward and backward values yields smoothed probabilities using the entire sequence.
A disciplined method
Set betaT(i)=1; recurse betat(i)=sum_j aij bj(o(t+1)) beta(t+1,j). Multiply alpha and beta, then normalise for each time.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether future evidence is available and appropriate to use. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A later outdoor reading increases the smoothed probability that the middle hidden state was Sunny compared with filtering at that moment.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Transfer to real-time monitoring where future data do not yet exist; distinguish filtered alerts from retrospective smoothing.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Smoothing can revise history but cannot remove model misspecification or look-ahead leakage in evaluation.
Evidence to retain: alpha and beta tables, normalisers, timestamps and evaluation protocol
Common misconception: A smoothed estimate would have been available to a live decision maker.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 8. Decode a path with Viterbi
The idea
Viterbi replaces sums with maxima and stores backpointers to recover the single most probable state path. It solves a different problem from marginal state estimation.
A disciplined method
Initialise delta, recurse with max over predecessor states, store the argmax, then backtrack from the best final state.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether the output needs one coherent path or marginal uncertainty. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
Two time points can each have the most likely marginal state Rainy while the best joint path includes Sunny because transitions reward coherence.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Change transition persistence and observe when Viterbi prefers fewer switches even with the same emissions.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: The best path can have low posterior mass when many near-tied paths exist.
Evidence to retain: delta table, backpointers, decoded path and path probability relative to total likelihood
Common misconception: Selecting the highest marginal state at each time always gives the Viterbi path.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 9. Learn parameters with Baum-Welch carefully
The idea
Baum-Welch is an expectation-maximisation procedure. It uses forward-backward expectations of state occupancy and transitions, then re-estimates parameters to increase likelihood.
A disciplined method
Choose initial parameters, run scaled E and M steps, track likelihood, use multiple starts and retain constraints or priors justified by the domain.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which local optimum and state labelling the fit represents. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
Two random starts fit similar likelihoods but swap state labels. External measurements help interpret which learned state is Degraded.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Transfer to a small dataset and compare maximum-likelihood learning with informative priors or supervised labels.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: EM can converge to local optima, degenerate emissions or uninterpretable states; likelihood can improve while usefulness worsens.
Evidence to retain: initialisations, likelihood trace, convergence rule, held-out results and interpretation checks
Common misconception: Baum-Welch discovers the true hidden states automatically.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 10. Prevent underflow and implementation errors
The idea
Sequence probabilities shrink multiplicatively. Direct floating-point computation can become zero even when the mathematical probability is positive.
A disciplined method
Use per-time scaling or log-space log-sum-exp, test against a short hand calculation, and verify that filtering probabilities normalise at every step.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is how scale factors are stored and converted back to log likelihood. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A 10,000-step sequence returns zero without scaling but a finite log likelihood with documented scale factors.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Change numeric precision and library implementation; compare likelihoods and decoded paths within tolerances.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Stable arithmetic prevents numeric failure but not statistical misspecification.
Evidence to retain: unit tests, seed, library version, scale factors and tolerance policy
Common misconception: More decimal places fix underflow without changing the algorithm.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 11. Diagnose state count, duration and residual dependence
The idea
Model selection must balance fit, interpretability and predictive performance. Standard HMM dwell times are geometric, and residual serial dependence signals missing structure.
A disciplined method
Compare held-out likelihood and task metrics across state counts, inspect dwell-time distributions, posterior uncertainty and residual autocorrelation, and test simpler baselines.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether an added state captures a real regime or only flexible noise. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A three-state model splits Degraded into two indistinguishable labels while residual durations remain too long, suggesting a duration model rather than another state.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Transfer to a process with scheduled phases and add time-varying transitions or a hidden semi-Markov model.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Information criteria depend on regularity assumptions and do not replace domain validation.
Evidence to retain: train-test split, selection criteria, dwell plots, residual tests and stability across starts
Common misconception: The model with the highest training likelihood is the best model.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 12. Test transfer, drift and decision consequences
The idea
An HMM fitted in one environment may fail when emissions, transitions, prevalence or sensors change. Decisions based on states require consequences and error costs to be explicit.
A disciplined method
Validate on a changed site or period, compare parameter and calibration drift, evaluate state uncertainty, and map false state assignments to real decision costs.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which parameters are assumed transportable. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A machine model moves to a hotter factory. Temperature emissions shift while transition dynamics partly persist; recalibration beats blind reuse.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Change the decision threshold and cost of a missed Degraded state; compare monitoring policies without changing the state model.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Good sequence likelihood does not guarantee fair, safe or useful decisions, especially for people.
Evidence to retain: external validation, drift dashboard, subgroup checks, decision costs and retraining trigger
Common misconception: A mathematically valid HMM can be deployed unchanged wherever the variable names match.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Field workbook
Build the two-state worked example in a spreadsheet or short script. Verify one joint path, the forward likelihood, one smoothed probability and the Viterbi path by hand before changing parameters. Then rerun after doubling state persistence and explain every changed result.
- Write the decision question in one sentence and name the decision owner.
- List observations separately from interpretations, with dates and provenance.
- Record at least two plausible alternatives before choosing a preferred account.
- Run the changed-context test and note which assumptions fail.
- State the limit, confidence and next review trigger in plain language.
| Question | Evidence | Failure signal |
|---|---|---|
| Are states defined? | External interpretation check | Labels swap or drift |
| Is inference correct? | Hand and unit tests | Likelihood mismatch |
| Do assumptions fit? | Duration and residual plots | Long memory remains |
| Does it transfer? | External validation | Calibration or emission shift |
Continue through the verified ecosystem
- Conditional random fields — discriminative sequence labels
- Markov random fields — undirected dependence and message passing
- Bayesian networks — directed conditional independence
- Partially observable Markov decision processes — beliefs linked to actions
- Robust Markov decision processes — policies under transition ambiguity
Frequently asked questions
What is hidden in an HMM?
The state sequence is latent. Observations are measured and probabilistically linked to each current state.
What does the forward algorithm compute?
It sums the probabilities of all state paths consistent with the observations, organised by ending state.
Why use the backward algorithm?
Together with forward values it incorporates future observations to compute smoothed state probabilities.
Is Viterbi the same as choosing the most likely state each time?
No. Viterbi finds the most probable joint path, whose transitions can make it differ from pointwise marginal choices.
Does Baum-Welch find the true parameters?
Not necessarily. It can converge to local optima and state labels may be non-identifiable without external interpretation.
When is a standard HMM a poor fit?
When duration is not geometric, memory extends beyond one state, transitions vary with covariates, or emissions retain dependence after conditioning.
Authoritative sources and further reading
- Rabiner, A Tutorial on Hidden Markov Models — classic algorithms and applications
- Proceedings of the IEEE DOI for Rabiner 1989 — canonical publication record
- Jurafsky and Martin, Hidden Markov Models — current academic chapter with worked algorithms
- Bilmes, A Gentle Tutorial of the EM Algorithm — EM and HMM parameter learning
- IEEE Xplore record — bibliographic and publication details
This guide explains a statistical model. Model outputs are conditional on data, assumptions and implementation; high-stakes uses require domain validation, human review and explicit error costs.