Numerical Reasoning Test 2026: How to Beat the Clock

Extraction comes before arithmetic

Every item in this format has three stages: find the figures the question needs, decide which operation answers it, then run the operation. Candidates rehearse the third stage and lose on the first two.

SHL's own candidate report says so in as many words. Its numerical test "measures the ability to make correct decisions or inferences from numerical data," and then this: "The emphasis in these tasks is on understanding and evaluating data rather than on computation."

So the preparation that helps is reading exhibits fast and correctly. Before touching a number, answer three questions about the item.

1. Which rows and columns does the question actually name? A five column table with one relevant column is the normal case, and the other four are there to cost you time.
2. What is being asked for, in what unit? A percentage, an absolute amount, a rate, a difference between two rates. Write the unit down before you compute.
3. Which single operation gets there? If you cannot name it in one phrase, you have misread the question rather than found a hard one.

Run those three and the arithmetic takes seconds. Skip them and you compute something correct that answers a different question, which scores the same as an error.

This is also why a calculator changes less than candidates expect. Where one is allowed, it removes the part that was never the constraint and leaves the extraction and the choice of operation exactly as they were. Where one is banned, the arithmetic on the page is sized to be done by hand. Either way the exhibit is the work.

What a numerical reasoning test measures, and what it does not

The same SHL report describes the data and tasks as "relevant to a business environment," and says people who score well "tend to have the capacity to understand numerical data and interpret mathematical information correctly." That is the working skill a firm screens for. A consultant is handed a client exhibit and has to say what it shows, quickly, in front of someone who will check.

What it does not measure is mathematical technique. There is no algebra, no calculus, and no memorized formula beyond percentages, ratios and averages. A candidate who revises mathematics for this is preparing for a test nobody is setting.

The length and shape come from the product. No general standard governs them. SHL's fact sheet for Verify Interactive numerical reasoning states an "Allowed Time (minutes) 18 minutes," a "Maximum Number of Questions 10 questions," "Number of Sittings One," and a "Question Format Interactive / Activity-based, Adaptive." It is published at shl.com, checked 29 September 2026.

Ten items in eighteen minutes changes the arithmetic of the sitting. Each item is worth a tenth of the outcome, so one careless extraction costs far more here than on a forty item paper.

Four operations cover almost every item you will meet: a percentage change between two figures, a share of a total, a rate applied to a quantity, and a weighted average. The weighted average is the one candidates get wrong, because the instinct is to average the averages. Two offices with margins of 20 and 30 percent do not give a combined margin of 25 percent unless they are the same size, and an item that hands you two sizes is asking for exactly that distinction.

Worked item one: percentage growth is not absolute growth

Question: which region grew fastest?

The trap is that the question does not say fastest in what. Read it as percentage growth, which is what "fastest" means for a rate, and the answer is East.

• North: 462 minus 420 is 42. 42 divided by 420 is 0.10, so 10.0 percent.
• South: 361 minus 380 is negative 19. Negative 19 divided by 380 is negative 0.05, so a 5.0 percent decline.
• East: 280 minus 250 is 30. 30 divided by 250 is 0.12, so 12.0 percent.

East grew fastest at 12.0 percent. North added the most revenue in absolute terms, 42 against East's 30, and a candidate who scans for the biggest jump picks North. Both figures are correct. Only one answers the question that was asked.

The habit that prevents this: before computing, write down whether the answer is a percentage or an amount. Two words on scrap paper, and the most common trap in the format stops working on you.

Worked item two: the profit held and the margin fell

Revenue rose from 80 million to 88 million. Costs rose from 60 million to 68 million.

Question: what happened to profit, and to margin?

Profit first. 80 minus 60 is 20. Then 88 minus 68 is 20. Profit is unchanged at 20 million.

Margin second. 20 divided by 80 is 0.25, so 25.0 percent. Then 20 divided by 88 is 0.2273, so 22.7 percent. Margin fell by 2.3 percentage points.

Both answers come from the same four numbers, and they point in different directions. That is the whole point of the item. A question asking whether the business improved has no single answer here, which is why the wording matters: profit held, margin fell, and revenue grew.

Note the unit in the margin answer. The fall is 2.3 percentage points, not 2.3 percent. As a percentage of the original margin the fall is 2.3 divided by 25, which is 9.2 percent. Both are true statements about the same movement, and an answer option list will often contain one of each.

Worked item three: two routes to the same number

A plant produces 1,200 units per day, six days a week. Four percent of output is scrap. Each good unit sells for 35.

Question: what is weekly revenue from good units?

First route. 1,200 times 6 is 7,200 units a week. Good units are 96 percent of that: 7,200 times 0.96 is 6,912. Revenue is 6,912 times 35, which is 241,920.

Second route. Gross revenue would be 7,200 times 35, which is 252,000. Scrap removes 4 percent: 252,000 times 0.04 is 10,080. Then 252,000 minus 10,080 is 241,920.

The two routes agree, and that agreement is the check worth building into your habit. On a timed test you will not have room to compute every item twice, so use the second route as an order of magnitude test rather than a full recomputation. Roughly 7,000 units at roughly 35 is roughly 245,000, so an answer near 240,000 is plausible and an answer near 24,000 or 2.4 million is a misplaced decimal.

Decimal placement is the single most common arithmetic error on this format, and a one second sanity check catches almost all of it.

The sanity check has a second use. When the answer options are spread widely, rounding is enough to pick one: if the options are 24,000, 241,920 and 2.4 million, the rough figure decides it without the exact calculation. When the options are close together, the spread itself tells you the item is testing precision, and that is the signal to do the full arithmetic carefully rather than fast. Reading the options before computing costs two seconds and sometimes saves the whole item.

Where the clock goes on a numerical reasoning test

Time in this format is spent in four places, and only one of them is arithmetic.

• Reading the exhibit. Do this once, properly, including the units in the column headers and any footnote. A footnote that says figures are in thousands has decided several answers before you start.
• Rereading the exhibit. This is the avoidable one, and it is caused by reading it badly the first time.
• Deciding the operation. Fast when the question is parsed correctly, slow when it is not.
• Computing. Seconds, on almost every item.

The practical budget on a ten item, eighteen minute test is one minute forty-eight per item, since 18 divided by 10 is 1.8 minutes. Spend the first twenty seconds on the exhibit and the question, name the operation out loud in your head, then compute. If you pass your budget on an item, commit to an answer and move rather than trading two later items for this one.

The skip rule depends on the platform

Almost every guide on this subject tells candidates to skip a hard item and come back to it. That advice is wrong on at least one platform in common use, and the publisher says so directly.

TestGorilla's candidate guide states: "Once skipped, you cannot go back and answer a question. We do not recommend skipping questions. Skipped questions count as wrong answers. If you don't know the answer, guess." It is published at candidates.testgorilla.com, checked 29 September 2026.

On that platform, a blank and a wrong answer score identically, so a guess is free and a skip is a loss. The behavior is not universal. Adaptive products, including the SHL one above, select the next item from your answers, which is a different mechanism again and one reason difficulty tends to rise as you do well. Finding it hard partway through is a signal about your performance rather than a problem.

The rule to carry: read the instruction screen before the first item, and let it override any general advice, including this page. Publishers put their own rules there, and the screen in front of you is current in a way that no guide can be.

One more platform difference worth knowing. Some products allow an employer to grant extra time as an accommodation, and some present practice items at the start whose answers are discarded. Where practice items exist, use them to learn the interface rather than to warm up the arithmetic, because an interface surprise on item one costs more than a slow first calculation.

The four labels to put on a wrong answer

A percentage score tells you nothing about what to fix. Sort every item you get wrong into one of four labels, and the label decides the correction.

1. Misread the question. You answered a different question correctly. Correction: write the unit down before computing.
2. Wrong extraction. Right method, wrong figures off the exhibit. Correction: read the headers and footnotes first, every time.
3. Wrong method. You picked an operation that does not answer the question. Correction: name the operation in one phrase before starting.
4. Ran out of time. Correction: a per item budget, and the discipline to commit and move.

After twenty items one label will dominate, and that label is your entire preparation. Most candidates who feel weak at mathematics discover they are mostly label one or label two, which is a reading problem with an arithmetic reputation.

The labels also tell you what to stop doing. A candidate whose errors are label three has no use for more mental arithmetic drills, and a candidate whose errors are label four has no use for more untimed papers. Both are common, and both waste the week before the sitting.

Keep the log in two columns, the item and the label, and nothing else. A longer log does not get written after the fourth set, and the count per label is the only output you need from it.

How it is scored, and what the employer keeps to itself

SHL's candidate report prints two things per test: a percentile against a comparison group that is named on the report, and a letter grade on a five point scale shown as E, D, C, B, A. On the sample report SHL publishes, numerical results are compared with a "General Population 2006" comparison group, so the benchmark is a dated population rather than an absolute bar.

SHL describes its cognitive range at shl.com/products/assessments as "automated, remotely proctored, and ask fewer questions thanks to our unique partial-scoring methodology." Partial scoring is why a ten item test can still separate candidates.

The report also states that "test results remain valid for about 12 to 18 months," which makes a second sitting inside the same cycle unlikely to be offered. It is worth reading that line twice before a first attempt taken casually, because the result may follow your application for longer than the season you are applying in.

What is not published anywhere is the cutoff. The publisher reports a percentile and a band; the employer decides which band continues, and no firm publishes that decision. Nor does anyone publish which publisher a given firm uses, because it varies by office and by cycle. Your invitation email names the platform, and it is the only source on that which is current by definition.

Practicing the reading rather than the sums

SHL publishes example questions and a full length practice test on its own candidate site, covering numerical, verbal, inductive, calculation and checking formats among others. It is the publisher's material, it is free, and it is where to start before paying for a question bank.

Work timed from the first session. Untimed practice trains the arithmetic, which is the part you already have, and leaves the pacing untouched. After each set, label every wrong answer with one of the four labels above and fix the dominant label before doing more volume.

We do not have a dedicated numerical reasoning drill, and claiming one would be easy and false. The free drills train the case version of the same skill, where a number has to be produced under a clock with the working shown, and how an answer is assessed sets out the weights they score against, read from the code that does the scoring. Our assessment simulators cover firm built formats rather than publisher tests, and none of the six is live yet, which that page states itself.

For the publisher mechanics, SHL style assessments covers the adaptive behavior and the personality questionnaire. For the other half of most screens, verbal reasoning covers the one rule that decides that format.