Why Tail Direction Comes Before the Formula
Many hypothesis-testing mistakes happen before any arithmetic begins. A candidate sees a sample mean, a standard error, and a significance level, then immediately reaches for a formula. But the formula cannot tell you which tail matters.
The first question is conceptual:
What would count as evidence against the null hypothesis?
If only unusually high outcomes matter, the test is right-tailed. If only unusually low outcomes matter, the test is left-tailed. If unusually high or unusually low outcomes both matter, the test is two-tailed.
This map keeps the test from becoming a pile of disconnected symbols.
The Null Is the Baseline, Not the Desired Result
The null hypothesis is the statement tested against the evidence. It commonly contains equality: equal to, at least, at most, no change, no difference, or no abnormal performance.
The alternative hypothesis is what the analyst is looking for evidence to support. It may be directional or non-directional.
Suppose Alder Fund claims its active strategy generates more than a 7.0% annual expected return. A researcher wants to test whether the evidence supports a return greater than 7.0%.
- Null hypothesis:
mu <= 7.0% - Alternative hypothesis:
mu > 7.0%
That is a right-tailed test because only unusually high sample results support the alternative.
Now suppose the researcher only wants to know whether the expected return differs from 7.0%, in either direction.
- Null hypothesis:
mu = 7.0% - Alternative hypothesis:
mu != 7.0%
That is a two-tailed test because unusually high and unusually low results both contradict the null.
One-Tailed Tests Put Alpha in One Direction
In a one-tailed test, the entire significance level sits in one tail.
If alpha = 5% and the test is right-tailed, the rejection region is the most extreme 5% on the right side of the null distribution.
Example:
Marlowe Advisors tests whether a bond-screening model has a mean monthly excess return greater than 0.20%.
H0: mu <= 0.20%Ha: mu > 0.20%- Significance level: 5%
- Test statistic: 1.82
- Right-tail critical value: 1.645
Because 1.82 > 1.645, reject the null. The evidence is far enough into the right tail to support the claim that mean excess return is greater than 0.20%.
Two-Tailed Tests Split Alpha Across Both Tails
A two-tailed test asks whether a parameter is different from the null value, not specifically higher or lower. Therefore the rejection probability is split across both tails.
For a 5% two-tailed test, each tail gets 2.5%.
Suppose Lydian Pension Consultants tests whether a manager's mean tracking difference is different from zero.
H0: mu = 0Ha: mu != 0- Significance level: 5%
- Test statistic: -2.12
- Two-tailed critical values: -1.96 and +1.96
Because -2.12 < -1.96, reject the null. The statistic is far enough into the left tail. A positive statistic above +1.96 would also reject the null because the test is two-tailed.
P-Values and Critical Values Must Use the Same Tail
The p-value approach and the critical-value approach should agree when they are based on the same alternative hypothesis.
For a right-tailed test:
- Critical-value approach: reject if the test statistic is greater than the right-tail critical value.
- P-value approach: reject if the right-tail p-value is less than alpha.
For a two-tailed test:
- Critical-value approach: reject if the statistic is beyond either critical boundary.
- P-value approach: reject if the two-tailed p-value is less than alpha.
The common mistake is mixing a one-tailed p-value with a two-tailed alpha, or choosing a right-tail critical value when the alternative is left-tailed.
Worked Example: Same Statistic, Different Alternative
Riverton Equity Research studies whether a new valuation screen changes monthly active return relative to zero. The test statistic from the sample is t = 1.74.
Case 1: The analyst predicted improvement.
H0: mu <= 0Ha: mu > 0- Right-tailed test at 5%
- Critical value: 1.645
Decision: reject, because 1.74 > 1.645.
Case 2: The analyst only predicted a difference.
H0: mu = 0Ha: mu != 0- Two-tailed test at 5%
- Critical values: -1.96 and +1.96
Decision: fail to reject, because 1.74 is not beyond +1.96.
The same statistic can lead to a different decision because the alternative hypothesis defines where the rejection region lives.
Exam Framing
On CFA exam questions, do not start by asking which formula appears familiar. Start with a small decision tree:
Then compute the statistic. Then compare it to the rejection region.
That order makes hypothesis testing less mysterious. The null is the baseline. The alternative points to the tail. The rejection region tells you how unusual the sample must be before you reject the null.