AICc for Short Time Series

Why Correct AIC for a Short Series?

A time-series model can fit a short observed series well simply because it has enough parameters to adapt to its particular ups and downs. The ordinary Akaike information criterion (AIC) penalizes model complexity, but in small samples its penalty can be insufficient: the apparent fit may be more optimistic than the fit the model would achieve on new observations. AICc adds a correction that is especially important when the number of fitted parameters is not small relative to the number of observations.

AICc is a criterion for comparing candidate models, not a test of whether a model is true. A smaller value is preferred among models fitted to the same response data under comparable likelihood conventions. Model adequacy still matters: a low criterion does not excuse residual patterns that indicate the model has missed important time-series structure.

Definition: For a model fitted by maximum likelihood, let $\ell$ be its maximized log-likelihood, $k$ the number of estimated parameters counted under a consistent convention, and $n$ the number of observations contributing to the likelihood. Then $$ \operatorname{AIC}=-2\ell+2k $$ and the small-sample corrected Akaike information criterion is $$ \operatorname{AICc}=\operatorname{AIC}+\frac{2k(k+1)}{n-k-1}, \qquad n>k+1. $$ The added fraction is the small-sample correction. If $n\leq k+1$, this formula is undefined and must not be used to rank the model.

For time-series models, $k$ typically includes estimated autoregressive and moving-average coefficients, any estimated mean or drift terms, and the innovation variance if it is estimated. Include other estimated parameters as well, such as seasonal coefficients when present. Use the same counting convention for every candidate. The value of $n$ is the number of observations actually used by the likelihood, not the forecast horizon. If likelihoods use different observations, initial-value treatments, or incompatible conditional and exact likelihood conventions, the resulting criteria may not be comparable.

Calculating and Interpreting the Correction

The correction depends on both model complexity and sample size. When $n$ is large compared with $k$, it is relatively small. When $n$ is only a little larger than $k+1$, its denominator is small and the correction can be substantial. This is the intended warning: a flexible model fitted to very little data should face a larger penalty than ordinary AIC alone imposes.

Worked Example: AIC Prefers the Larger Model, but AICc Does Not

A hypothetical analyst fits two candidate models to the same 20 quarterly observations using comparable maximum-likelihood fits. Model A has $k=2$ estimated parameters and maximized log-likelihood $\ell=-30$. Model B has $k=4$ and $\ell=-27.5$. Calculate both criteria and select the preferred model by each.

Method and parameter check. Use $n=20$, and use the given total parameter counts, which include all estimated quantities under the analyst’s common convention. Both models satisfy $n>k+1$: $20>3$ and $20>5$.

Model A. First calculate AIC: $$ \operatorname{AIC}_A=-2(-30)+2(2)=60+4=64. $$ Its correction is $$ \frac{2(2)(2+1)}{20-2-1} =\frac{12}{17} =0.705882\ldots. $$ Therefore, $\operatorname{AICc}_A=64.705882\ldots$, or 64.706 to three decimals.

Model B. Its AIC is $$ \operatorname{AIC}_B=-2(-27.5)+2(4)=55+8=63. $$ Its correction is $$ \frac{2(4)(4+1)}{20-4-1} =\frac{40}{15} =2.666667\ldots. $$ Therefore, $\operatorname{AICc}_B=63+2.666667\ldots=65.666667\ldots$, or 65.667 to three decimals.

AIC prefers Model B because $63<64$, while AICc prefers Model A because $64.706<65.667$. The fit improvement for Model B is not enough to overcome the larger small-sample correction. Common mistake avoided: choosing the model with the smaller AIC and assuming that the AICc ranking must be identical.

The reason for adding a correction is tied to what AIC estimates. AIC estimates a model’s expected information loss when it is used beyond the data on which it was fitted. Fitting parameters to a particular sample improves the in-sample likelihood partly by adapting to random features of that sample. This makes in-sample fit optimistic as an estimate of future fit. The $2k$ penalty is the usual first-order adjustment for that optimism; in common regular likelihood settings, the small-sample correction accounts for additional finite-sample bias. It is a standard correction, not an exact universal identity for every possible time-series likelihood.

Why the Penalty Grows with Complexity

Write the additional AICc penalty as $c(k)=2k(k+1)/(n-k-1)$, keeping $n$ fixed. For successive admissible parameter counts, the difference is

c(k+1)-c(k) = \frac{2(k+1)(2n-k-2)} {(n-k-2)(n-k-1)} >0, \qquad n>k+2.

Both denominator factors are positive under the stated condition. The numerator is also positive, so adding one parameter increases the correction. The denominator also shrinks as $k$ approaches $n-1$, which helps explain why the correction can rise sharply near the boundary. For fixed $k$, the correction tends toward zero as $n$ grows; for fixed $n$, it increases with $k$. These properties are useful checks on a calculation: a more complex model should not receive a smaller correction solely because it has more parameters.

Worked Example: The Same Models at Two Sample Sizes

Two candidate time-series models have AIC values of 42 and 39. The first has $k=2$, and the second has $k=5$. Compare their AICc values for a hypothetical series with $n=10$ observations, then repeat the correction calculation for $n=50$. In each case, assume the models use the same observed series and comparable likelihoods.

At $n=10$. The first model satisfies $10>2+1$. Its correction is $$ \frac{2(2)(3)}{10-2-1}=\frac{12}{7}=1.714286\ldots, $$ so its AICc is $42+1.714286\ldots=\mathbf{43.714}$ to three decimals.

The second model also satisfies the requirement, since $10>5+1$, but its denominator is much smaller: $$ \frac{2(5)(6)}{10-5-1}=\frac{60}{4}=15. $$ Its AICc is $39+15=\mathbf{54.000}$. AIC selects the second model, but AICc selects the first: $43.714<54.000$.

At $n=50$. For the first model the correction is $12/(50-2-1)=12/47=0.255319\ldots$, giving AICc $42.255$. For the second it is $60/(50-5-1)=60/44=1.363636\ldots$, giving AICc $40.364$. At this larger sample size, AICc selects the second model. The added penalty has become smaller relative to the AIC difference. All AICc values here are rounded to three decimals only at the final step. Common mistake avoided: treating the correction as a fixed charge for a particular model, rather than recalculating it using the actual likelihood sample size.

AICc as a Model-Comparison Procedure

A careful comparison requires more than substituting values in a formula. First establish that the candidates address the same series and that their likelihoods can be compared. Then count each model’s estimated parameters consistently, identify the correct likelihood sample size, and check $n>k+1$ separately for every candidate. Only after these checks should you calculate AICc and compare the resulting values.

Check comparability.
Confirm that the models are fitted to the same response observations and use comparable likelihood conventions. A criterion difference is not meaningful if it mainly reflects different data or likelihood definitions.

Set $n$ and count $k$.
Use the observations entering the likelihood. Count all estimated model parameters, including an estimated innovation variance, consistently across candidates.

Check the formula’s domain.
Verify $n>k+1$ for each model. If the condition fails, do not report a finite AICc from this formula.

Calculate and compare.
Compute $-2\ell+2k$, add $2k(k+1)/(n-k-1)$, and prefer the smaller AICc among the comparable candidates.

Worked Example: Selecting Among Three Candidates

An analyst fits three hypothetical models to the same 16 monthly observations, with comparable likelihoods. Their parameter counts and AIC values are: Model P, $k=2$, AIC $=48$; Model Q, $k=3$, AIC $=46$; and Model R, $k=5$, AIC $=44$. Determine the AICc ranking.

Check admissibility. The largest parameter count is 5, and $16>5+1$, so all three candidates satisfy the denominator condition.

Model P: its correction is $2(2)(3)/(16-2-1)=12/13=0.923077\ldots$. Its AICc is $48.923077\ldots$, or 48.923 to three decimals.

Model Q: its correction is $2(3)(4)/(16-3-1)=24/12=2$. Its AICc is $46+2=\mathbf{48.000}$.

Model R: its correction is $2(5)(6)/(16-5-1)=60/10=6$. Its AICc is $44+6=\mathbf{50.000}$.

The AIC ranking is R, Q, P, from smallest to largest. The AICc ranking is Q, P, R. Thus Model Q is preferred by AICc, despite not having the smallest AIC. The calculation also illustrates that a strong enough likelihood improvement can still compensate for the correction: AICc does not automatically reject every more complex model. Common mistake avoided: subtracting the correction from AIC. It is added, so that extra complexity receives an additional penalty.

Limits and Common Exam Traps

AICc does not replace time-series model checking. A candidate with the smallest AICc may still leave serial structure in its residuals, or may have a coefficient pattern that makes the fitted model questionable. Use the criterion as one part of model selection, alongside the earlier material on likelihood comparability and residual diagnostics. Also remember that a very large correction does not prove that a model is unsuitable in every respect; it signals that the criterion is strongly penalizing its complexity relative to the available sample.

Using the wrong denominator. The denominator is $n-k-1$, not $n-k$. Check the subtraction before calculating.
Subtracting instead of adding. AICc is AIC plus the correction. Subtraction rewards, rather than penalizes, the small-sample complexity adjustment.
Omitting an estimated parameter. If the innovation variance is estimated, it counts under the usual total-parameter convention. Apply one consistent convention to every candidate.
Using the series length without checking the likelihood. Some fitting procedures do not use every observed value in the same way. Use the actual number of observations entering the likelihood being compared.
Applying the formula outside its domain. When $n\leq k+1$, the denominator is zero or negative. Do not interpret the resulting value as a valid AICc.

Key takeaway: AICc adds $2k(k+1)/(n-k-1)$ to AIC. The correction becomes more consequential as model complexity rises or the likelihood sample becomes short. Count parameters and observations consistently, check $n>k+1$, and compare AICc only for models with comparable likelihoods.

Check Your Understanding

For each problem, use the stated parameter counts and assume the candidate likelihoods are comparable unless the problem says otherwise. Round numerical AICc values to three decimals.

A model has $n=14$, $k=2$, and AIC $=30$. Calculate its correction and AICc. (Answer: correction $=12/11=1.091$; AICc $=31.091$.)
A model has $n=25$, $k=4$, and AIC $=72$. Is the formula admissible, and what is its AICc? (Answer: Yes, $25>5$; correction $=40/20=2$; AICc $=74.000$.)
Two models fitted to the same 12 observations have $(k,\operatorname{AIC})=(2,40)$ and $(4,37)$. Calculate both AICc values and select by AICc. (Answer: $41.333$ and $42.714$, respectively; select the first model.)
A candidate has $n=7$, $k=6$, and AIC $=20$. Can its AICc be calculated by the stated formula? (Answer: No; $n-k-1=0$, so the formula is undefined.)
Two candidates have $(k,\operatorname{AIC})=(2,60)$ and $(5,57)$, based on $n=30$ observations. Calculate their AICc values and choose the preferred candidate. (Answer: $60+12/27=60.444$ and $57+60/24=59.500$; select the second candidate.)

What You'll Learn