MATH7970 Chap.8 Estimation and Model Diagnostics
Estimation and Model Diagnostics
Numbers for a candidate
Moments and likelihood estimate a specified model; they do not validate its assumptions. Convergence, starting values and parameter boundaries deserve attention.
Residuals return the verdict
Residual time plots, ACFs, distribution checks and squared-residual diagnostics identify unmodelled structure.
Information criteria balance fit and complexity but cannot replace those checks.
Estimation inherits specification
Method of moments matches theoretical and sample relationships such as means and autocovariances. Maximum likelihood evaluates a joint distribution for the observed sequence. Neither rescues a poor order, missing seasonality or changing variance.
Their estimates can differ in finite samples, and both depend on initial treatment and distributional assumptions that should be recorded with the result.
Audit the optimiser
Likelihood surfaces may be flat, multimodal or close to stability and invertibility boundaries. Inspect convergence messages, try alternative sensible starting values and compare with simpler estimates.
A reported maximum is weak evidence when nearby starts produce materially different parameters. Confirm that the resulting coefficients define an admissible process before using standard errors, long-run means or forecasts.
Give each diagnostic its own target
The residual time plot reveals breaks, outliers and changing spread. Residual ACF tests remaining linear mean dependence.
Distribution checks matter for likelihood inference and forecast intervals. Squared-residual ACF tests conditional variance dependence. Passing one target does not imply the others pass, so record each separately and connect every failure to the assumption it challenges.
Revise one feature at a time
Information criteria compare fit and complexity inside a candidate set but do not certify adequacy.
When a diagnostic fails, name the lag, sign or variance pattern and make the smallest plausible repair. Refit and repeat the same check. Changing several orders together obscures which component mattered and encourages search over noise.
Preserve a validation period so the final specification must also forecast data it did not shape.
Separate estimation uncertainty from innovation uncertainty
Standard forecast formulas often condition on estimated parameters as if known, while finite samples leave those estimates uncertain. The distinction matters most with short series, persistent roots and elaborate seasonal structures.
Compare coefficient uncertainty across candidates and avoid treating a barely identified extra term as a precise mechanism. Bootstrap or other extended procedures may incorporate parameter uncertainty, but even without them the written interpretation should note when a narrow model-based interval omits instability visible in estimates or across reasonable specifications.
What this chapter covers
- 01
method of moments
- 02
maximum likelihood
- 03
numerical optimisation
- 04
residual ACF
- 05
information criteria
- 06
squared residuals
Worked example · free
Diagnose seasonal residuals
- 2Name the evidence.
- 2Identify the failed assumption.
- 2Propose a targeted repair.
Key terms
- Method Of Moments
- Estimation by matching theoretical and sample moments.
- Maximum Likelihood
- Estimation by maximising the probability model assigned to observed data.
- Residual
- The observed value minus the fitted conditional mean.
- Information Criterion
- A fit measure penalised for model complexity.
- Diagnostic Checking
- Testing fitted residuals for structure the model should have removed.
Estimation and Model Diagnostics FAQ
Does the largest likelihood always win?
Only comparable models fitted to the same data can be contrasted, and complexity, diagnostics and forecasting still matter. An optimiser can also settle at a poor local solution or boundary.
What do squared residuals reveal?
They expose dependence in conditional variance that may be absent from the raw residual ACF. Persistent spikes motivate an ARCH-family variance specification after the mean is cleaned.
Does a clean residual plot prove the model true?
No. It means the selected diagnostics found no material remaining pattern at the lags and features examined. A new holdout period or structural break can still overturn that judgement.
Exam move
Use a four-column diagnostic log: evidence, failed assumption, smallest repair and repeated check. This keeps model revision tied to observable failure. Keep a diagnostic notebook where every failed check leads to one targeted revision. Record starting values and convergence results for likelihood estimates, then compare a second start.
For each candidate, list residual mean, notable ACF lags, squared-residual pattern, criterion values and validation error. Practise rejecting the lowest criterion when residual structure remains, and retaining a simpler model when an added coefficient is unstable. This makes specification search auditable rather than a hunt for significant terms. End every diagnostic entry with evidence that would reverse the decision.
This keeps a retained model provisional and makes the next validation period an explicit test rather than ceremonial confirmation. Keep rejected candidates in the notebook so repeated searches do not rediscover them.
Working through Estimation and Model Diagnostics in MATH7970? Sia is AskSia’s AI Mathematics tutor — ask any MATH7970 Estimation and Model Diagnostics question and get a clear, step-by-step explanation grounded in how MATH7970 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
