Card Decision

What you see

Four face-down decks labelled A, B, C and D sit in a row, each with its own card-back mark so they can be told apart without colour. Under each deck is a count of how many times you have drawn from it. You draw by clicking a deck or pressing the keys 1 to 4.

Under the decks, a bar shows the net result of the card you just drew, growing left from a centre mark at zero for a loss or right for a gain, with the net figure beside it. A line under the board says what the card paid, for example Deck B won 100 and lost 250. Two pills above the board show the round you are on (Round 13 / 30) and your running total in points. After you check your answer, each deck also shows what it alone would have paid over the whole run.

One run of the game is one trial: a fixed number of draws, 30 at level 1 rising to 90 at level 10. You can draw from any deck at any time and switch as often as you like. There is no time limit, and speed has no effect on the score.

How the decks are built

The decks follow the design of the original Iowa task. It helps to know the design, because it is exactly what you are trying to discover.

• Two decks pay a bigger win on every card (100 at level 1, rising to 150 at level 10), but their losses are larger still. Over time they lose money.
• Two decks pay a smaller win on every card (50 at level 1, rising to 65), with losses small enough that they come out ahead.
• In each pair, one deck loses often and a little (five loss cards in every ten) and the other loses rarely and a lot (one loss card in every ten).
• Every block of ten cards from a deck nets exactly its designed amount. Where the losses fall inside each block is shuffled, so you cannot predict which card will hurt.
• Each deck deals its own cards in order. Your seventh draw from a deck is that deck's seventh card, whatever you drew from the other decks in between.
• Which letter holds which deck is drawn afresh every run. Deck A can be a paying deck in one run and a losing one in the next, so a letter tells you nothing; only the cards do.

What changes with level

Three things tighten as you climb: the runs get longer, the gap between good and bad decks shrinks (from 25 a card to 10), and the temptation grows, because the losing decks pay ever-bigger wins. At level 10, a losing deck pays 150 on every card, well over twice the good decks' 65.

What it measures

Card Decision feeds the Learning from feedback facet and carries 38 per cent of your Decision Making score, the largest single share.

The task descends from the Iowa Gambling Task published by Antoine Bechara, Antonio Damasio, Hanna Damasio and Steven Anderson in 1994. In the original, participants received a $2,000 loan of facsimile money and made 100 selections from four decks. Decks A and B paid $100 a card, C and D paid $50. After ten cards from deck A, a participant had won $1,000 but paid $1,250 in penalties, a net loss of $250; deck B lost the same amount through rarer, larger penalties. Decks C and D netted $250 per ten cards. Healthy controls drifted towards C and D; patients with ventromedial prefrontal damage kept choosing A and B, as if guided only by immediate prospects. Our decks keep the same shape and scale the numbers by level, but unlike the original they are dealt to the letters in a fresh order every run.

What it asks of you is a chain of skills: sampling enough to form a view, keeping a running tally per deck, noticing that a deck's average is negative despite frequent wins, and switching behaviour to match. In reinforcement-learning terms, you are estimating each deck's value from prediction errors and choosing by those estimates.

The task has a known quirk. Many healthy people prefer the rare-loss decks, including the losing one. This is known as the "prominent deck B" phenomenon, which Lin and colleagues confirmed in 2007, and a 2013 analysis of eight data sets by Steingroever and colleagues found healthy participants often prefer decks with infrequent losses. The reason is the one you met in the subject lessons: when you learn from experience, a rare loss has less impact than it deserves. Nine quiet +100 cards feel like a great deck. Knowing this bias is half of beating it.

How it is scored here

Your score for a run is value captured: your realised total divided by what the best single deck would have earned over the same number of draws. Because every run is a whole number of ten-card blocks and each good deck nets exactly its designed amount per block, that benchmark is simply the good decks' net per card times the number of draws: 750 at level 1 (25 × 30), 900 at level 10 (10 × 90).

• The ratio is clipped to the range 0 to 100 per cent. A negative total scores 0.
• 98 per cent or more counts as correct and moves you up one level for the next run. Zero counts as incorrect and moves you down one. Anything in between holds your level. New sessions start at level 3.
• Time is not scored at all.

Three consequences follow straight from this formula.

1. Every draw from a losing deck costs about 200 ÷ draws points on average. A bad draw loses the good-deck net and adds the bad-deck net, so it costs twice the per-card edge. At level 1 (30 draws) that is about 7 points a draw; at level 10 (90 draws) about 2.
2. One rare big loss can wipe out a run. At level 1, the big-win rare-loss deck's bad card nets −1,150, more than the whole 750 benchmark. Sampling that deck is a gamble on missing its one bad card in ten.
3. Swings can only hurt you. Good luck above the benchmark is capped at 100 per cent; bad luck is not. Alternating between the two good decks leaves each with part-finished blocks and adds swings. In our simulations of the scoring, splitting draws evenly between the good decks averaged about 84 to 92 per cent at levels 1, 5 and 10, against exactly 100 for staying with one good deck.

Your session report adds descriptive measures that never change the score: Draws from the good decks, Draws until 7 of 10 came from the good decks (the first point where seven of ten consecutive draws came from good decks), Change in good-deck draws, first third to last third, Draws away from your best-looking deck (model estimate), and How strongly each result moved your choices (model estimate), a learning rate fitted to your choices with the prediction-error rule from the subject lessons. Good decks here means the two paying decks of that run, whichever letters they had.

Across sessions your Card Decision score sits on our 0–100 scale, where 50 means what the difficulty you work at predicts. Harder runs earn more credit for the same value captured. A score appears after about four runs, small samples are pulled towards 50, and recent sessions count for more than old ones.

Where you will meet it

Candidate-preparation sites describe a Cards game in the pymetrics battery with four decks in which each draw wins or loses money, the same design family as this task. Pymetrics was used by Unilever as a first screen for graduate hiring, according to Business Insider. Prep sites also describe an Energy game in the Arctic Shores battery in which you learn which generators add power and which drain it, a close cousin. These descriptions are candidate-reported rather than vendor documentation.

Outside hiring, the Iowa Gambling Task remains a standard clinical instrument for decision-making problems after brain injury and in addiction research, which is why so much is known about how people approach it.

The method

1. Sample each deck briefly

   Draw one or two cards from each deck in turn. That is four to eight draws. Note each deck's win size and any losses. Do not judge yet.

2. Use the win size as a warning

   The decks that pay the larger win on every card are the ones built to lose. Bigger win per card goes with bigger losses in this design. If you have already seen a big-win deck produce a large loss, you have your confirmation.

3. Price by the block, not the card

   For any deck, ask: over ten cards, what do the wins add up to, and what have the losses cost so far? A small-win deck with a −50 now and then is fine; a big-win deck with a −250 every other card is not.

4. Commit to one good deck

   Pick one small-win deck and draw from it for the rest of the run. Staying with a single deck means its blocks complete cleanly and its designed profit lands in full.

5. Expect its losses and hold

   Your good deck will produce losses too, either small ones often or one sizeable one per ten. That is built in. Do not switch because of a single loss; switch only if a deck's running tally turns clearly negative.

6. No exploring in the last third

   Any draw from a losing deck late in the run is pure cost: there is no time left to use what you learn.

Worked example 1: which deck after eight draws?

You are at level 1 and have drawn each deck twice. Your notes from this run:

Show the answer

Commit to A or D, and pick one. C has the best tally so far, but it pays the big win of 100, which is the warning sign: big-win decks are built to lose. Its rare loss has simply not arrived yet. At level 1 it is a −1,250 penalty that turns that card into −1,150, and it comes once in every ten cards. B's frequent −250 losses are already showing. A and D pay only 50, yet they net +25 a card over each ten-card block.

If you choose D, expect one −250 loss (a −200 card) per ten cards. If you choose A, expect five −50 losses per ten, each wiping out that card's win. Both earn exactly 250 per block. Stay with the one you choose for the remaining 22 draws.

Worked example 2: the good deck bites

Level 6, 60 draws. You committed to the small-win rare-loss deck after your opening samples. Twelve draws in, it deals +55 with −390, a net −335 on one card. Your running total drops sharply. Should you move to the other small-win deck?

Show the answer

Stay. That −390 penalty is this deck's one loss in its current block of ten. Over those ten cards it still nets 10 × 55 − 390 = +160, which is exactly the level's designed 16 per card. The other good deck earns the same +160 per block, just through five losses of 78.

Switching would not raise your expected earnings at all. It would leave this deck's next block partly drawn and start a partial block in the other, adding swings that the capped score cannot reward. And if you had seen this card as a reason to try a big-win deck, you would be paying about 3 points of score per draw at this level, plus the risk of a −1,240 card.

Worked example 3: what a ten-card experiment costs

At level 1, a candidate spends the first ten draws on the frequent-loss big-win deck "to be sure", then switches to a good deck for the remaining 20.

Show the answer

Those ten cards form one complete block, which nets exactly −250. The next 20 draws from one good deck form two complete blocks at +250 each, +500. The total is +250 against a benchmark of 750: 33 per cent. The same ten-draw experiment at level 10 costs less (it lands at 78 per cent) because the edge per card is smaller and the run is longer, but it is still the biggest single mistake you can make. Two or three samples of a deck tell you its win size; ten tell you nothing more that you need.

Tips and traps

Levelling up

As levels rise, the edge per card shrinks from 25 to 10, the losing decks' wins grow from 100 to 150, and runs lengthen from 30 to 90 draws. Two adjustments help.

• Sampling gets cheaper per draw but not safer. With 90 draws, one bad draw costs only about 2 points on average, but a single rare loss (−1,450 net at level 10) is still bigger than the whole 900-point benchmark. Sample the rare-loss big-win deck at most once, or not at all once you have seen its win size.
• Your patience has to grow. The good decks' own losses get bigger relative to their wins (at level 10 the rare-loss good deck charges 550 once per ten cards of 65). Trust the block arithmetic, not the feeling after a bad card.

Because moving up needs 98 per cent of the benchmark, expect to hold a level for several runs while you tighten your sampling. That is normal; the score already adjusts for the difficulty you are working at.

1. Week 1
   • Three runs; sample each deck twice, then commit
   • After each run, check Draws until 7 of 10 came from the good decks
   • Write the per-deck tally at the moment you committed
2. Week 2
   • Three runs; cut sampling to one draw per deck
   • Stay with a single good deck all run, whatever its losses
   • Check that Change in good-deck draws, first third to last third is positive
3. Week 3
   • Two runs at your current level; aim for 98 per cent or more
   • Compare your fitted learning rate across runs: aim for steady, not jumpy
   • Revisit worked example 2 if you switched after a loss

Check yourself

5 practice questions

1. At level 1 the big-win rare-loss deck pays +100 on nine cards and has one −1,250 penalty per ten. What does it average per card?
2. How is a run of Card Decision scored?
3. You have committed to a good deck. It deals a large loss. What should you usually do?

and 2 more

Create a free account to answerSign in

Sources

1. Bechara, A., Damasio, A. R., Damasio, H., & Anderson, S. W. (1994). Insensitivity to future consequences following damage to human prefrontal cortex. Cognition, 50(1–3), 7–15. doi.org/10.1016/0010-0277(94)90018-3
2. Bechara, A., Damasio, H., Tranel, D., & Damasio, A. R. (1997). Deciding advantageously before knowing the advantageous strategy. Science, 275(5304), 1293–1295. doi.org/10.1126/science.275.5304.1293
3. Maia, T. V., & McClelland, J. L. (2004). A reexamination of the evidence for the somatic marker hypothesis: What participants really know in the Iowa gambling task. PNAS, 101(45), 16075–16080. doi.org/10.1073/pnas.0406666101
4. Lin, C.-H., Chiu, Y.-C., Lee, P.-L., & Hsieh, J.-C. (2007). Is deck B a disadvantageous deck in the Iowa Gambling Task? Behavioral and Brain Functions, 3, 16. doi.org/10.1186/1744-9081-3-16
5. Steingroever, H., Wetzels, R., Horstmann, A., Neumann, J., & Wagenmakers, E.-J. (2013). Performance of healthy participants on the Iowa Gambling Task. Psychological Assessment, 25(1), 180–193. doi.org/10.1037/a0029929
6. Hertwig, R., & Erev, I. (2009). The description–experience gap in risky choice. Trends in Cognitive Sciences, 13(12), 517–523. doi.org/10.1016/j.tics.2009.09.004
7. Buelow, M. T., & Suhr, J. A. (2009). Construct validity of the Iowa Gambling Task. Neuropsychology Review, 19(1), 102–114. doi.org/10.1007/s11065-009-9083-4
8. JobTestPrep. Free pymetrics games online practice and guide (candidate-preparation site; descriptions are unofficial). jobtestprep.com
9. JobTestPrep UK. Arctic Shores assessment (candidate-preparation site; descriptions are unofficial). jobtestprep.co.uk
10. Feloni, R. (2017, June 28). Consumer-goods giant Unilever has been hiring employees using brain games and artificial intelligence. Business Insider. businessinsider.com

Put it into practice: Card DecisionSign in to try it