Skip to content

Formal hypothesis tests

Translation and update of §"Contrastes Formales de Hipótesis" of Treadway (2011). Full references are in docs/PROVENANCE.md §9; "Guerrero (2026)" below is the unpublished manuscript on hybrid seasonal models, which is where the per-frequency critical values come from. The tests are the same; what is new is that the critical values the manual left as an empty table now exist, and that this document says which ones are tabulated, which are simulated and which are interpolated.


1. The rule that comes before every test

Treadway states it first, and it is not a preamble:

A formal test should be applied only on a model that is efficiently estimated, statistically adequate and parsimoniously parameterised, and should be used to simplify the parameterisation only when the message of the test is very clear — not when the null is accepted or rejected marginally at one conventional level.

From an over-parameterised model one can obtain any answer; parameterise enough and one obtains the answer one wants.

The operational form of that rule: a formal test on an inadequate model is not a weak test — it is not a test at all. ADF, KPSS and the rest are initial specification tools; the tests below belong at the end of the process, on a model that has passed diagnosis.

fue computes none of these automatically. Each is a likelihood ratio between two runs of the program: fit the model free, fit it again with the parameter fixed at the boundary, and compare logelf in the two .out files. That is why every section below shows the .inp fragment that imposes the restriction — a coefficient line whose flag is 0 instead of 1.

Not covered here: Jarque-Bera and Ljung-Box, whose statistics fue already prints in the .out.


2. Non-stationarity of an AR(1) factor — Shin-Fuller

H₀: φ₁ = 1 H₁: φ₁ < 1

             ⎧ 2[ l(φ̂₁) − l(φ̄₁) ]     if φ̂₁ ≤ φ̄₁
  SF   =     ⎨
             ⎩ 0                       if φ̂₁ > φ̄₁

  φ̄₁ = 1 − 4/n

with n the number of observations of the potentially under-differenced model under H₁, and φ₁ the parameter of the largest AR(1) factor in the model. l(φ̂₁) is logelf from the unrestricted fit; l(φ̄₁) requires re-estimating with φ₁ fixed at φ̄₁ — for n = 100, at 0.96:

** Number and orders of regular AR operators:
1 1
**
.96 0

Shin & Fuller (1998) show the asymptotics do not differ between an ARMA(p+1,q) with a strictly stationary and invertible ARMA(p,q) factor and a plain AR(1), and give evidence that the test dominates the alternatives in size and power.

Critical values — Shin & Fuller (1998), Table II. Reject H₀ when the statistic exceeds them:

n 10% 5% 1%
25 1.02 1.68 3.33
50 1.06 1.75 3.41
100 1.07 1.75 3.41
250 1.07 1.76 3.44
500 1.08 1.77 3.46

The manual printed this table empty. These are the published values, and art's formal_tests.shin_fuller interpolates them linearly between the five sample sizes — an approximation for intermediate n, not a tabulation.

The statistic is Φ̂₁ᵤ = L_free − L_constrained, not 2·ΔL. The manual writes the 2[·] form; art implements the paper's Φ̂₁ᵤ, and the critical values above belong to Φ̂₁ᵤ. Using one convention with the other's table doubles or halves the statistic.


3. Non-invertibility of an MA — DCD

For MA(1) and MA_f operators estimated inside the invertibility region, the generalised likelihood ratio of Davis & Dunsmuir and Davis, Chen & Dunsmuir.

Regular MA(1), positive parameter. H₀: θ₁ = 1, H₁: θ₁ < 1

  DCD = 2[ l(θ̂₁) − l(θ₁ = 1) ]
** Number and orders of regular MA operators:
1 1
**
1 0

Frequency-fixed MA_f. H₀: λ_f = −1, H₁: λ_f > −1

  DCD = 2[ l(λ̂_f) − l(λ_f = −1) ]

The same fragment applies, with the parameter fixed at −1.

Critical values, and which regime a frequency belongs to

This is where the update matters most, because the law is not the same at every frequency. It is governed by the order of the factor, not by which frequency it sits at — Monte Carlo evidence in Guerrero (2026); implemented in art.formal_tests:

regime which frequencies d.o.f. pile-up 10% 5% 1%
real root (s=1) regular MA(1), f = 0, Nyquist f = s/2 1 0.6575 1.00 1.94 4.41
complex pair (s=2) interior f = 1 … s/2−1 2 0.616 1.11 2.04 4.52

The real-root values are ≈ n-invariant. The complex-pair values carry a mild finite-sample dependence, tabulated by simulation and interpolated between these sample sizes:

n 10% 5% 1%
120 1.12 2.06 4.64
240 1.13 2.07 4.52
480 1.10 2.04 4.53
960 1.11 2.03 4.52
→ ∞ 1.11 2.04 4.52

These supersede the interpolated values of the thesis (1.07 / 2.02 / 4.52).

The published simulations, quoted

The tables above are what art implements; Guerrero (2026, Table 1) publishes the whole simulation, with Monte Carlo standard errors, and the two are close enough that the interpolation costs little — which is worth knowing before anyone re-runs it:

regime n pile-up 10% 5% 1%
real root (f = 0, 6) 120 0.651 (.002) 1.00 (.02) 1.95 (.03) 4.42 (.09)
240 0.645 (.002) 1.03 (.02) 1.97 (.03) 4.47 (.07)
480 0.651 (.003) 1.01 (.02) 1.91 (.03) 4.38 (.09)
960 0.650 (.003) 1.02 (.02) 1.96 (.05) 4.52 (.11)
complex pair (f = 1…5) 120 0.615 (.002) 1.12 (.01) 2.06 (.02) 4.64 (.04)
240 0.615 (.002) 1.13 (.01) 2.07 (.02) 4.52 (.05)
480 0.617 (.002) 1.10 (.01) 2.04 (.02) 4.53 (.06)
960 0.617 (.002) 1.11 (.02) 2.03 (.03) 4.52 (.07)
Davis MA(1) 0.6575 1.00 1.94 4.41
thesis interpolation 1.07 2.02 4.52

And the realistic case (Guerrero 2026, Table 2) — the witness together with a mean and the nine surviving harmonics, k = 10 regressors — which is the one to use on a real model:

n pile-up 10% 5% 1%
120 0.559 (.002) 1.63 (.02) 2.87 (.04) 5.81 (.10)
240 0.586 (.002) 1.34 (.02) 2.39 (.03) 5.16 (.11)
480 0.601 (.002) 1.23 (.02) 2.20 (.03) 4.67 (.08)
960 0.611 (.003) 1.13 (.02) 2.10 (.04) 4.70 (.10)
bare, n→∞ 0.616 1.11 2.04 4.52

The gap between the two tables is the whole point: at n = 120 the 5% value moves from 2.06 to 2.87 once the model carries what a real model carries.

The seasonal AR_f statistic

The MEG sweep of §5 needs the other half — the analogue of Shin-Fuller at a seasonal frequency — and Guerrero (2026, Table 3) tabulates it, for the statistic Φ̂_f with fixed truncation ρ_m = 1 − c/n, 2×10⁴ replications:

factor deterministic s c 10% 5% 1%
interior f = 1…5 harmonic (2-dim) 2 3 1.34 2.12 3.90
[1.32, 1.37] [2.09, 2.15] [3.85, 3.94]
real (f = 0, Nyquist) constant / alternator 1 4 1.06 1.75 3.47
[1.05, 1.06] [1.70, 1.79] [3.46, 3.48]
Shin-Fuller Table II, AR(1), c=4 1.07 1.75 3.41

Two things to take from it. The real-frequency row reproduces Shin-Fuller's Table II — as it must, since it is the same law — which is a check on the simulation rather than a new result. And the interior frequencies use c = 3, not 4: the truncation constant of the two-dimensional case is not the one of the AR(1), and using ρ_m = 1 − 4/n there is testing a different point.

art does not implement this statistic yet: formal_tests carries the DCD regimes and Shin-Fuller, and the MEG sweep decides on the MA_f witness. The table is quoted here because the values exist and the sweep is where they belong.

Two production caveats, both measured, both easy to get wrong.

  1. Compute the boundary likelihood exactly. The restricted fit must profile over a fixed grid rather than let the free MA optimiser walk to the boundary: the optimiser is biased at the second-order non-invertibility boundary and produces a spurious pile-up of ≈0.82 against the correct ≈0.62.

  2. In a realistic model — one carrying a mean and deterministic harmonics — the correct finite-sample critical values are HIGHER: at n = 120, 1.63 / 2.87 / 5.81 at 10/5/1%, an effect that vanishes as n grows. Using the bare values there over-rejects.


4. Fixed frequency for an AR(2) with imaginary roots

Worth testing when the estimated frequency of an AR(2) factor is close to one of f = 1 … s/2−1 relative to its standard error.

H₀: f = 1 (say) H₁: f ≠ 1

  LR = 2[ l(H₁) − l(H₀) ]     ~ χ²₁ under H₀

l(H₁) from the free AR(2) with imaginary roots, l(H₀) from the same model with the factor constrained to the fixed frequency (an AR_f operator). Reject at 1−α when the statistic exceeds the χ²₁ quantile.

This one is an ordinary LR with a standard distribution, because the restriction is interior — unlike §2 and §3, which sit on a boundary and therefore have non-standard laws.


5. Stochastic seasonality — the MEG sweep

Starting from a model with fully deterministic seasonality, evaluate frequency by frequency whether the seasonality is stochastic instead.

Two ordering rules, and they matter:

  • Evaluate stochastic seasonality BEFORE simplifying the deterministic seasonal terms, so the deterministic representation is as flexible as possible while you test.
  • Simplify the intervention terms and the ARMA structure FIRST, so the model is as parsimonious as possible when you start.

For each frequency f = 1 … s/2, estimate one model that adds both:

  • the homogeneously non-stationary AR_f operator at that frequency, and
  • an MA_f operator at the same frequency as an over-differencing witness (seed λ_f at, say, −0.9).

Applying the non-stationary AR_f annihilates the deterministic seasonal terms at that frequency, which is what makes the comparison clean. For monthly data that is six models, one per frequency, each with its DCD test on the MA_f.

Decision: if the MA_f comes out literally non-invertible — or the DCD says λ̂_f does not differ significantly from the non-invertible value — reject stochastic seasonality at that frequency.

Do not apply this rigidly, and Treadway is explicit:

Even when the DCD tests say an MA_f is invertible, with and without the corresponding AR_f overfit, it is sometimes worth integrating at that frequency anyway, to see whether the representation loses anything noticeable. Faced with ambiguity — the data not discriminating between stochastic and deterministic — the analyst should choose the representation that suits a pre-established criterion.

No program and no algorithm can replace the analyst's judgement in building models from data.

There is a companion to this in the confirmatory direction: at f = 0 the Shin-Fuller test (§2) and the DCD test (§3) have opposite nulls and bracket the quasi-cancellation band, so their disagreement is itself the diagnostic. See Guerrero (2026), the comparison table.


6. Simplification of intervention terms

When an intervention carries more than one parameter, test the hypothesis of zero long-run gain with a Student t.

  g ≡ ω_s(1) / δ_r(1)

the long-run effect on the output of a permanent unit increase in the input. Imposing g = 0 on two consecutive steps, for example, is the same as specifying that intervention as a single impulse in the level.

The point is not only saving parameters: it is that the simplified form is the one that says what actually happened. An intervention you can interpret against the extra-sample information about the incident is worth more than one that fits marginally better.


7. Where these are implemented

fue computes none of them automatically — it provides logelf and the parameter table, and the tests are ratios between runs.

test implemented in notes
Shin-Fuller (§2) art.formal_tests.shin_fuller Φ̂₁ᵤ convention, Table II interpolated
DCD, DCD_f (§3) art.formal_tests.dcd, dcd_overdiff_regular regime chosen by factor order
MEG sweep (§5) art.formal_tests.meg sweep or explicit frequency list
fixed frequency (§4) LR by hand; χ²₁
gain t-test (§6) drtran reports the gain and its s.e.

The division is deliberate and is the architecture of the suite: fue is the engine, art is the criterion. A likelihood ratio needs two fits and a decision rule; the engine supplies the fits and stays out of the decision.